Foreign Object Detection in Railway Tracks: A Survey

Foreign Object Detection in Railway Tracks: A Survey

Aadhrika A K

B Tech Scholar Department of Information Technology, Government Engineering College Idukki

Hannah Liji

B Tech Scholar Department of Information Technology Government Engineering College Idukki

Divya Prasad K H

Assistant Professor, Department of Information Technology Government Engineering College Idukki

Aarcha A

B Tech Scholar, Department of Information Technology, Government Engineering College Idukki

Keerthana A

B Tech Scholar Department of Information Technology, Government Engineering College Idukki

Abstract – Railway transportation is highly vulnerable to accidents caused by foreign objects obstructing tracks, which can endanger passenger safety and cause significant delays, damage to infrastructure, and interruptions in train services. Conventional monitoring methods are inefficient due to their dependence on manual supervision and poor adaptability in adverse conditions. This paper presents a survey of foreign object detection techniques in railway tracks using computer vision and deep learning approaches. In the proposed system, images captured under adverse conditions are first enhanced using image enhancement techniques and then passed to the YOLO-based object detection module, whereas normal images are directly processed for object detection. Once an obstruction is detected, automated alerts containing object type, image, risk level, location, and time are transmitted through a mobile application, supported by a backend for incident logging and rapid communication. By integrating computer vision and real-time notification systems, the solution offers a reliable and scalable approach to improving railway safety and ensuring uninterrupted train services

1. INTRODUCTION

   Railway tracks are vital for transportation, but they are vulnerable to foreign objects such as animals, debris, or unauthorized persons, which can cause accidents and operational delays. Manual monitoring of tracks is time- consuming and prone to human error, so automated detection systems are necessary to ensure safety and efficiency.

   Existing railway monitoring systems often rely on human inspections or drone-based surveillance. While drones can provide comprehensive coverage, they may be costly to deploy. For demonstration purposes, many recent studies use video input from cameras uses video input from cameras to extract frames and analyze them with

   AI models, detecting and classifying potential threats on the tracks. The system can also assess the risk level of detected objects and generate alerts for timely action, helping to maintain safe and uninterrupted railway operations.

   Implementing this detection system enables railway authorities to enhance safety by minimizing the risk of accidents caused by foreign objects. It also improves operational efficiency by reducing delays and preventing potential damage to infrastructure. By leveraging AI-driven analysis, the system provides accurate, real-time monitoring and timely alerts, representing a significant advancement toward smarter, safer, and more reliable railway operations.

2. RELATED WORK
   1. Obstacle Detection System for Railways using IoT Sensors

      Early works in railway foreign object detection focused on cost- effective sensing technologies. Ahmad Akmal Aziz, Khairul Anuar Mohamad, and Afishah Alias [1] demonstrated how IoT- enabled hardware could be integrated with rail infrastructure to create a continuous monitoring system. By combining sensor data collection, wireless communication, and cloud visualization, the system showcased the potential of real- time alerts in preventing accidents. However, the limitations of short detection range and environmental susceptibility highlighted the need for more advanced visual and AI-based methods in later research.

      The authors implemented an IoT-based prototype using ultrasonic sensors placed near the railway tracks. Sensor signals were processed by an Arduino NodeMCU microcontroller and

      transmitted via Wi-Fi to the FavorIoT platform. A radar-style interface was developed to visualize detected obstacles in real time, emphasizing low power consumption, cost efficiency, and simple integration with existing railway monitoring infrastructure. Experiments showed reliable detection of static obstacles at close range with minimal latency, but performance dropped in noisy outdoor environments and under adverse weather. The study concluded that while IoT- based monitoring is a valuable first layer of defense, future systems must combine sensor networks with AI-driven computer vision for improved robustness and scalability.

   2. Image Processing Based Level Crossing Detection and Foreign Objects Recognition Approach in Railways

      Building on classical computer vision techniques, Tastimur, Karakose, and Akn [2] explored image processing pipelines to detect foreign objects at railway level crossings. Their study demonstrated how color space transformations, edge detection, and feature extraction could be effectively combined to identify potential hazards in real time. While providing a computationally lightweight solution, the approach highlighted the limitations of fixed-threshold methods and challenged researchers to adopt more adaptive and learning-based techniques in subsequent works.

      The methodology employed a sequential pipeline including YCbCr color transformations, Prewitt edge detection, Gaussian filtering, and Hough transforms to delineate critical regions along the tracks. Foreign object detection was performed using HSV-based segmentation and connected component analysis. Tastimur et al. [2] reported that while the system was efficient for clear daytime scenarios, its performance degraded under variable illumination, shadows, and partial occlusions. This study underscored the importance of incorporating more robust detection methods, paving the way for deep learning- based approaches in railway safety monitoring.

   3. Fast Detection Study of Foreign Object Intrusion on Railway Track

      In efforts to improve detection speed and reduce computational overhead, Niu and Hou [3] proposed a lightweight approach for rapid identification of foreign objects along railway tracks. This study emphasized real-time processing, recognizing that high-speed train environments demand fast and reliable detection methods. Building on traditional image processing pipelines, the method integrated multi-frame analysis and historical background modeling, offering a practical compromise between simplicity and responsiveness. The work highlighted the growing awareness that detection frameworks must not only be accurate but also feasible for continuous monitoring in operational railway settings.

      Technically, the approach defined Railway Track Dangerous Areas (DAS) using Canny edge detection and Hough transforms, followed by multi-background modeling

      and pixel-level decision-making. By analyzing differences across multiple frames and historical data, the system could detect anomalies quickly, achieving high processing speed with minimal computational resources. While the model performed well in controlled or simple environments, its effectiveness decreased under complex lighting conditions, occlusions, or small-object scenarios, demonstrating the need for subsequent integration of more robust AI-based detection mechanisms.

   4. Dense Attention Railway Foreign Object Detection Algorthm Based on Mask R-CNN

      As deep learning began to dominate railway safety research, Gao introduced advanced convolutional neural networks to improve detection robustness, particularly in challenging environmental conditions. The study highlighted the limitations of traditional computer vision pipelines and proposed a Mask R- CNN-based framework enhanced with dense attention mechanisms. By focusing on multi-scale feature extraction and attention-guided learning, the method achieved significantly higher detection accuracy for both visible and infrared images, marking a shift toward high-performance deep learning approaches in railway monitoring systems.

      Gaos [4] O-Mask R-CNN integrated Dense Feature Pyramid Networks (FPN) with Convolutional Block Attention Modules (CBAM) to capture fine-grained features, especially under weak illumination and infrared conditions. The Region Proposal Network (RPN) was further optimized using IoU-based K- means clustering, enhancing anchor box adaptability. The model achieved detection accuracy of 98.26% on infrared images and 98.85% on normal images with fast convergence. While demonstrating exceptional precision, the approach required substantial computational resources and showed reduced stability under severe weather conditions, emphasizing the trade- off between high accuracy and operational efficiency in practical deployments.

   5. A Novel Strategy of Two-stage Cascaded CNN and Overhaul Knowledge Distillation

      To address the growing demand for real-time railway foreign object detection while maintaining high accuracy, Meng et al. [5] proposed a two-stage cascaded CNN framework. This study represented a significant evolution from single-stage detectors by combining a lightweight preliminary classifier with a high- precision detector, reducing computational overhead without sacrificing detection quality. The approach illustrated how cascading networks could filter irrelevant images first, allowing the more complex model to focus only on potential intrusions, demonstrating a practical strategy for balancing speed and precision in railway monitoring systems.

      In their implementation, the first stage employed a ResNet- tiny classifier to quickly categorize images as normal or

      intruded, thereby reducing unnecessary computation. Images

      flagged as intruded were passed to YOLOv3 for detailed foreign

      object detection. Additionally, the authors applied Overhaul Knowledge Distillation (OKD) to transfer knowledge from a large ResNet-50 teacher network to the smaller student network, improving classification performance. Experimental results showed 99.33% classification accuracy and a processing speed of 65.79 FPS, making it suitable for real- time monitoring. The study noted challenges in handling rare or complex intrusions, emphasizing the ongoing need to refine model generalization for diverse railway environments.

   6. Foreign Object Detection in Railway Images Based on an Efficient Two-Stage Framework

      Following the trend of multi-stage detection systems, Chen, Meng, and Jiang [6] proposed an efficient two-stage framework to improve both detection accuracy and computational efficiency in railway foreign object detection. This study built upon previous cascaded approaches by incorporating attention mechanisms and lightweight feature extractors, demonstrating how intelligent feature selection and multi-scale processing can enhance detection of small or partially occluded objects along railway tracks. The work illustrated the importance of optimizing the balance between detection precision and real-time applicability, reflecting a maturing phase in the evolution of AI-based railway safety systems.

      In the technical implementation, a lightweight classifier leveraging Inverted Residual Units, Selective Kernel Convolutions, and CBAM attention modules was used to filter incoming images, identifying potential intrusion candidates. The flagged images were then processed by YOLOv3 with a DarkNet-53 backbone and Feature Pyramid Network (FPN) to handle multi-scale object detection efficiently. The system achieved a classification accuracy of 96.85% and a detection mean Average Precision (mAP) of 85.89%, illustrating strong performance in detecting various foreign objects. However, the study was limited to only three object categories, suggesting that future work should explore scalability and robustness across more diverse and complex railway environments.

   7. Automatic Identification Method of Foreign Body Intrusion in Railway Transportation Track

      As research progressed, the focus shifted toward combining optimization algorithms with machine learning for more precise and automated foreign object detection. Niu [7] introduced a system that not only identified intrusions but also optimized the monitoring network itself, reflecting a trend toward intelligent, adaptive railway surveillance. This work demonstrated the potential of integrating multiple computational techniques to improve both detection performance and system efficiency in complex railway environments.

      Technically, Linjie Niu [7] implemented a hybrid framework combining improved LeaderRank algorithms, Gaussian mixture modeling, and CNNFirefly hybrid

      optimization. The system optimized surveillance node placement while enhancing segmentation and classification of detected objects. Experiments showed 93.2% accuracy and 92.1% recall, indicating strong detection capabilities. However, the method demanded significant CPU resources and careful parameter tuning, highlighting the trade-offs between accuracy, computational cost, and practical deployment in real-world railway operations.

   8. Research on Foreign Object Intrusion Detection for Railway Tracks Utilizing Risk Assessment and YOLO Detection

      Building upon earlier CNN- and YOLO-based methods, Ning et al. focused on integrating foreign object detection with risk assessment to provide not only accurate detection but also actionable safety insights. This approach reflected a shift in railway monitoring research toward intelligent decision-making frameworks that classify threats based on severity, allowing operators to prioritize responses and enhance operational safety. The study highlighted how lightweight models could be combined with attention mechanisms and track segmentation to improve both detection and assessment under real-world conditions.

      Shanping Ning [8] and colleagues implemented MobileNetV3-CATr as a lightweight backbone with channel attention and Transformer layers, combined with BiFPN-Lite for efficient multi-scale feature fusion. For track segmentation, SE- U- Net was employed, achieving 90.53% AMIoU. Detected objects were classified into risk categorieslow, medium, and highto support proactive safety measures. While the system outperformed baseline YOLO models in accuracy, it was slower than YOLOv5s and tested on a limited number of object classes, demonstrating that future work should focus on scaling the approach and improving inference speed for deployment on edge devices.

   9. Multi-block SSD Based on Small Object Detection for UAV Railway Scene Surveillance

      As railway foreign object detection research evolved, small- object detection became increasingly critical, particularly in UAV-based monitoring scenarios. Li, Dong, Li, Zhang, Zhang, and Xiao [9] emphasized that minor obstacles such as stones, small animals, or debris, though often overlooked by standard detectors, can pose serious safety risks. Their study highlighted how dividing images into smaller overlapping blocks and processing them separately could enhance detection accuracy while maintaining computational feasibility. This work marked a clear progression from conventional single-shot detection toward more refined, block-based strategies to handle complex railway environments.

      Yundong Li, Han Dong, Hongguang Li, Xueyan Zhang, Baochang Zhang, and Zhifeng iao [9] implemented a multi- block SSD (Single Shot MultiBox Detector) framework for UAV- captured railway images. The images were segmented into overlapping blocks processed individually, with results merged using region splicing and suppression techniques. Transfer learning was applied to adapt the model from general datasets to

      railway-specific imagery, resulting in a significant improvement in small-object detection recall from 56.4% to 92.1%. Despite its success, the approach incurred slower processing speeds and required extensive labeled data, illustrating the trade-off between recall, efficiency, and practical deployment constraints in real- world UAV monitoring systems.

   10. Detection Methods With Image Recognition for Specific Obstacles in the Urban Rail Area

       As railway safety research evolved, urban rail systems posed unique challenges due to higher traffic density, complex surroundings, and diverse obstacle types. Traditional sensors and basic image processing methods struggled to handle the variability and rapid changes in urban environments, prompting the adoption of more robust image recognition techniques. The focus shifted toward leveraging machine learning and deep learning models to detect specific obstacles such as pedestrians, vehicles, and construction materials, ensuring timely alerts for train operators.

       Shen et al. [10] implemented an image recognition- based framework that combined preprocessing techniques with deep convolutional networks to detect specific urban rail obstacles. The method included edge detection, segmentation, and feature extraction followed by classification using a CNN trained on a curated urban railway dataset. Their system demonstrated high accuracy in detecting stationary and moving obstacles, with improved adaptability under varying lighting and weather conditions. However, the approach required extensive labeled datasets and faced computational challenges for real-time inference, highlighting the need for future optimizations such as lightweight networks or multimodal integrations to enhance efficiency and scalability.

   11. Research on the Method of Foreign Object Detection for Railway Tracks Based on Deep Learning

       With the increasing demand for automated railway safety, Ning, Ding, and Chen [11] explored deep learning approaches to detect foreign objects on railway tracks. Their work built upon prior YOLO-based and CNN frameworks, aiming to improve detection speed and accuracy, especially under challenging conditions such as low lighting or partial occlusions. By integrating lightweight backbones, attention mechanisms, and segmentation modules, the system could precisely identify objects on tracks and classify them according to risk levels, providing actionable insights for railway safety management. This approach highlighted the transition from traditional image processing to modern AI- driven monitoring systems in railway operations.

       The authors implemented MobileNetV3 as the backbone network, enhanced with channel attention and BiFPN for multi- scale feature fusion. SE-U-Net segmentation ensured that only track-relevant regions were analyzed, reducing false positives from surroundings. The framework was tested on a custom railway dataset and demonstrated high precision and recall, outperforming baseline YOLO models in both

       detection accuracy and inference speed. Despite its strengths, the model required considerable computational resources for real- time deployment, and careful parameter tuning was necessary for optimal performance. This work illustrates the practical potential of combining deep learning with risk-based assessment to advance railway foreign object detection and safety monitoring.

   12. Research on Foreign Object Intrusion Detection in Railway Tracks based on MSL-YOLO

       With the increasing complexity of railway environments, traditional YOLO-based detection methods often struggle with small objects, occlusion, and varying lighting conditions. Researchers recognized the need for enhanced architectures capable of maintaining high detection accuracy while achieving real-time performance. Multi-scale feature fusion and attention mechanisms became a prominent solution to address these challenges.

       Hongxia Niu, Dingchao Feng, and Tao Hou [12] proposed MSL-YOLO, an improved YOLOv8n model integrating Multi- scale Shared Convolution Modules (MSCM), StarBlocks, and Efficient Multi-scale Attention (EMA) mechanisms to improve feature representation across scales. A Lightweight Shared Convolutional Detection Head (LSCD) with Dynamic Focus Loss further optimized bounding box regression. Evaluations demonstrated a 94.3% mAP at 277 FPS, balancing accuracy and speed. Despite these improvements, the systems performance decreased under extreme illumination or heavy occlusion, indicating future research directions in adaptive feature enhancement and robust small-object detection.

   13. Railway Foreign Object Intrusion Detection Using UAV Images and YOLO-UAT

       The use of UAVs for railway monitoring brought new opportunities for real-time, wide-area surveillance. Building on earlier YOLO-based detection systems, researchers focused on lightweight yet accurate models that could be deployed on edge devices like Jetson Nano for immediate inference. UAV images, with varying angles, altitudes, and lighting conditions, demanded models capable of handling occlusion, motion blur, and small- object detection without excessive computation.

       Yang Yang, Zhanhao Liu, Junming Chen, Haiming Gao, and Tao Wang proposed [13] YOLO-UAT, an optimized variant of YOLOv5s. They used EfficientNet-B0 as the backbone, incorporated CBAM attention modules, and applied K-means++ clustering for anchor optimization. The system achieved 91.5% mAP with 36% fewer parameters and ran at 26.4 FPS on Jetson Nano, demonstrating the feasibility of deploying powerful detectors on resource-constrained devices. The study highlighted UAV-based monitoring’s potential for dynamic railway surveillance, though challenges remained for night-time operations and adverse weather conditions.

   14. Improved YOLOv5 Based on Attention Mechanism and FasterNet

       Recent research emphasized combining lightweight convolutional strategies with attention mechanisms to improve detection efficiency without sacrificing accuracy. Building on YOLOv5 frameworks, this work targeted railway foreign object detection under diverse lighting and environmental conditions while keeping computational cost low for real-time deployment on edge devices. Integrating attention mechanisms allowed the model to focus on critical features, improving small-object and occluded-object detection.

       Zongqing Qi, Danqing Ma, Jingyu Xu, Ao Xiang, and Hedi Qu [14] proposed an enhanced YOLOv5 model trained on the large-scale AARFOD dataset containing 48,000 railway and aviation images. They incorporated Partial Convolutions from FasterNet and a Normalization-based Attention Module (NAM), which reduced parameters by 25% and FLOPs by over 10%, while achieving higher accuracy than the baseline YOLOv5. The experiments demonstrated that this approach effectively balances speed and detection performance, providing a practical solution for railway monitoring, particularly in environments with diverse object sizes, lighting, and weather conditions.

   15. YOLOv7-based Research on Foreign Object Intrusion Detection on Tracks

       The evolution of YOLO-based detectors has greatly influenced railway foreign object detection, with a continuous push for models that balance speed and accuracy. YOLOv7 emerged as a promising solution for real-time applications due to its optimized backbone and enhanced feature aggregation. Its capability to detect objects across varying scales made it suitable for complex railway environments where obstacles can appear in multiple sizes and positions.

       Xinnan Cai and Xuewen Ding [15] conductd a study applying YOLOv7 to railway track surveillance for foreign object intrusion detection. They trained the model on a custom dataset of 2,035 images, focusing on common railway obstructions such as stones, animals, and debris. The system achieved a precision of 0.816, recall of 0.667, and mAP@0.5 of 0.657, demonstrating competitive real-time detection performance compared to manual inspection methods. Despite its effectiveness, the study noted that YOLOv7s performance lagged behind newer YOLO variants like YOLOv8 and transformer-based approaches, highlighting the rapid evolution of deep learning models in railway safety monitoring.

   16. Transformer Based Foreign Object Detection Method in Railway Scenarios

       To leverage global context modeling for complex railway environments, transformer-based architectures were integrated into foreign object detection pipelines. This approach enabled the model to capture long-range dependencies and subtle spatial relationships, which traditional CNNs sometimes miss. The work demonstrated a shift toward hybrid CNN- transformer frameworks capable of enhancing detection accuracy in cluttered or low-light track

       scenarios while providing richer feature representation for decision-making.

       Cai and Ding [16] implemented a Swin Transformer-based YOLOX model, replacing PANet with BiFPN for enhanced multi- scale feature fusion. They further introduced Coordinate Attention modules to improve localization of elongated and small objects along the railway track. The model achieved 94.8% mAP, outperforming YOLOv5 and Faster R-CNN baselines. However, the high computational cost and slower inference speed limited deployment on resource-constrained platforms, indicating a trade- off between detection precision and real-time applicability in edge devices.

   17. Railway-CLIP: A Multimodal Model for Abnormal Object Detection

       As railway safety research progressed, purely visual models showed limitations in detecting novel or unusual objects not present in training datasets. To address this, multimodal approaches that combine visual and textual information began to emerge. These methods allowed models to generalize better and detect anomalies in a zero-shot manner, offering a significant advantage for real- world railway monitoring where unforeseen foreign objects may appear.

       Zhang et al. [17] developed Railway-CLIP, a multimodal framework combining CLIP, SAM, and Vision Transformers with Gaussian mixture modeling. The system aligned image features and textual prompts in a shared embedding space, enabling the detection of unseen anomalies such as suspended objects on catenary lines or irregular items along tracks. Evaluations showed 97.25% AUROC, demonstrating strong generalization. Despite its high accuracy, the method demanded substantial computational resources and complex integration of multiple AI components, making deployment challenging for real-time edge systems.

   18. Intelligent Railway Foreign Object Detection: A Semi- supervised Convolutional Autoencoder

       To overcome the persistent challenge of limited labeled datasets in railway foreign object detection, researchers explored semi-supervised learning approaches. These methods leverage abundant normal track images while minimizing reliance on rare abnormal samples, making the detection system more scalable and adaptable. This shift addressed a major bottleneck in practical railway monitoring where collecting comprehensive abnormal datasets is often infeasible.

       Tiange Wang, Zijun Zhang, Fangfang Yang, and Kwok- Leung Tsui [18] proposed a semi-supervised convolutional autoencoder trained exclusively on normal railway images. Using adversarial learning, their system detected anomalies through reconstruction errors and generated pixel-level localization maps for precise identification of foreign objects. Tested on Hong Kong MTR datasets, their model outperformed traditional approaches such as One-Class SVM and GAN-based methods, demonstrating higher sensitivity to rare anomalies while maintaining practical efficiency. Their study highlighted

       that semi-supervised learning can effectively bridge the gap between data scarcity and robust anomaly detection in railway environments.

   19. Track Foreign Object Image Augmentation using PLCA- pix2pixGAN

       To address the persistent challenge of limited railway foreign object datasets, researchers explored advanced data augmentation techniques. Generating realistic synthetic images allows models to learn from diverse scenarios without the need for costly real-world data collection. This approach helps improve the generalization and robustness of detection systems, particularly in handling rare or unusual object types.

       Xinyu Fan, Xuxu Yang, Feifei Hou, Cuipu Xi, and Yijun Wang [19] proposed the PLCA-pix2pixGAN framework for generating synthetic railway images containing foreign objects. The method overlays object templates, such as cows, cars, and sheep, onto real railway backgrounds under varied lighting and weather conditions. A channel-attention- guided generator ensures precise edge fidelity, producing high-quality synthetic images that preserve object realism. Experiments showed that models trained on these augmented datasets achieved higher SSIM and PSNR scores compared to conventional GAN-based augmentation, improving detection accuracy in complex scenarios. However, the methods reliance on predefined object templates limited generalization to unseen object types, highlighting areas for future improvement in generative augmentation for railway safety applications.

   20. RailFOD23: A Dataset for Foreign Object Detection on Transmission Lines

   A major limitation in railway foreign object detection research has been the scarcity of large, high-quality datasets. Without sufficient annotated images covering diverse object types, lighting, weather, and environmental variations, even state-of-the-art detection models struggle to generalize. Addressing this challenge, researchers have begun creating comprehensive datasets that combine real-world images with AI- generated imagery to boost both scale and diversity.

   Zhichao Chen, Jie Yang, Zhicheng Feng, and Hao Zhu

   [20] introduced RailFOD23, a large public dataset containing over 14,000 annotated images of foreign objects on

   transmission lines and railways. The dataset combines real photographs with AI-generated images using Stable Diffusion and ESRGAN techniques to enhance resolution and realism. Multiple object categories, such as plastic bags, balloons, and bird nests, are included, ensuring broad coverage of potential hazards. RailFOD23 enables benchmarking of modern object detection models like YOLOv8, which demonstrated strong accuracy and robustness on the dataset. By providing a standardized and diverse dataset, this work significantly advances the foundation for developing and evaluating reliable railway foreign object detection systems and addresses the long- standing issue of data scarcity in the domain.

3. COMPARATIVE ANALYSIS

   This section presents a comprehensive analysis of 20 influential research papers on Railway Foreign Object Detection (FOD) systems and surveillance technologies. It systematically compares their proposed models, core techniques, and primary focus areas, offering a structured overview of the field’s evolution. By organizing this information, the table provides valuable insights into current trends, deep learning methodologies, and key advancements in ensuring railway safety. It serves as a practical reference for researchers and practitioners aiming to design, evaluate, or implement robust, efficient, and real-time FOD solutions for uninterrupted train services.

   Table 1 summarizes the key techniques, models, and focus areas of major research works in foreign object detection in railway.

   Paper Title	Lead Authors	Proposed Model/Method	Core Innovation / Technique	Focus Area / Challenge
   Image processing based level crossing detection and foreign objects	C. Tatimur, M. Karaköse, and E. Akn	Vision-based Image Processing Method	Uses YCbCr, Hough transformation, and HSV color transformation for level crossing area detection and distance	Prevention of accidents at Level Crossings.

   recognition approach in railways			estimation.
   Automatic identification method of foreign body intrusion in railway transportation track based on improved LeaderRank identification of key points	Linjie Niu	Improved LeaderRank + Hybrid Deep Learning	Uses LeaderRank to identify optimal monitoring points; combined with improved Gaussian Mixture Model and hybrid DL model.	Foreign body intrusion in railway transportation track.
   Railway CLIP: A multimodal model for abnormal object detection in high-speed railway	J. Zhang, Q. Guan, J. Liu, Y.Huang and J. Guo	Railway- CLIP (Multimodal Model)	Improved LeaderRank + Hybrid Deep Learning	Uses LeaderRank to identify optimal monitoring points; combined with improved Gaussian Mixture Model and hybrid DL model.
   Research on foreign object intrusion detection in railway tracks based on MSL- YOLO	H.Niu, D. Feng, and T. Hou	MSL-YOLO (YOLO variant)	Multi-scale Shared Convolution Module (MSCM) and Efficient Multi-scale Attention (EMA) to enhance feature representation and efficiency.	Railway tracks; complex backgrounds, variable lighting, need for real-time, multi-scale detection.
   Foreign object detection in railway images based on an efficient two-stage convolutional neural network	W. Chen, S.Meng, and Y. Jiang	Efficient Two-Stage Framework	Stage 1 for image classification (intrusion/no intrusion), followed by Stage 2 for object detection.	General railway images; preventing train accidents from collisions.
   A novel strategy of two stage cascaded CNN and overhaul knowledge distillation for fast railway foreign objects intrusion detection	S. Meng, W. Chen, and Y. Jiang	Two-stage Cascaded CNN + Overhaul Knowledge Distillation (OKD)	Uses a cascaded CNN for efficient filtering combined with Overhaul Knowledge Distillation to speed up the process.	Achieving fast railway foreign objects intrusion detection (Efficiency).
   A dense attention railway foreign object detection algorithm based	Shuang Gao	O-Mask R- CNN (Optimized Mask R-	Integrates Densely connected FPN and Convolutional Attention Mechanism (CBAM) for	Detection in infrared and low illumination conditions.

   on Mask R-CNN		CNN)	low-contrast and small object recognition.
   Research on foreign object intrusion detection for railway tracks utilizing risk assessment and YOLO detection	S. Ning, R. Guo, P. Guo, L.Xiong and B. Chen	YOLO Detection + Risk Assessment Model	Combines object detection (YOLO) with a risk assessment model to quantify and distinguish the hazard level of the foreign object.	Prioritizing foreign objects based on potential hazard level to improve decision-making.
   Track foreign object image augmentation using PLCA- pix2pixGAN	X. Fan, Y. Zhang and H. Liu	PLCA- pix2pixGAN (GAN-based Data Augmentatio n)	Uses a Generative Adversarial Network (GAN) with Perceptual Loss and Channel Attention to generate high-quality synthetic images.	Data augmentation for detection systems constrained by small datasets and low sample diversity.
   Automatic railway traffic foreign object detection based on FFR- CNN	T.Ye, J.Zhao and L. Xu	Feature Fusion Refine NN (FFR-CNN)	An automatic object detection system using a Feature Fusion Refine network.	Railway Traffic Object Detection specifically under Shunting Mode (low speed < 45 km/h).
   Railway foreign object intrusion detection using UAV images and YOLO UAT	Y. Yang, Z. Liu, J. Chen, H. Gao and T. Wang	YOLO-UAT (Improved YOLOv5s)	Replaces YOLOv5s backbone with EfficientNet; uses C3_CBAM module (attention) and K- means++ clustering for anchors.	Foreign object intrusion detection using UAV images (aerial perspective).
   Improved YOLOv5 based on attention mechanism and FasterNet for foreign object detection on railway and airway tracks	Z. Qi, D. Ma, J.Xu, A.Xiang and H. Qu	Improved YOLOv5 (with FasterNet integration)	Integrates FasterNet and attention mechanisms; introduces the AARFOD (Aero and Rail Foreign Object Detection) dataset.	Foreign object detection on Railway AND Airport runways (Dual domain).
   YOLOv7-based research on foreign object intrusion detection on tracks	Xinnan Cai	YOLOv7	Application and training of the high-accuracy, real-time YOLOv7 model for track FOD.	Real-time and accurate detection of track foreign objects.
   Intelligent railway foreign object detection: A semi-	T. Wang, Z. Zhang, F. Yang and K.-	Semi supervised Convolutioal	Uses a Convolutional Autoencoder and reconstruction	Detection of foreign objects with unknown categories

   supervised convolutional autoencoder based method	L. Tsui	Autoencoder	discrimiator for anomaly detection without prior labels on foreign objects.	(Anomaly Detection).
   RailSegVITNet: Lightweight VIT- based segmentation for railway foreign object detection	Z. Chen, J. Yang, Z. Feng and H. Zhu	RailSegVITN et (Vision Transformer based)	Integrates lightweight bottleneck blocks and separable self-attention for semantic segmentation of the track surface.	Precise foreign object detection through accurate track segmentation.
   Multi-block SSD based on small object detection for UAV railway scene surveillance	Y. Li, H. Dong, H. Li, X. Zhang, B. Zhang and Z. Xiao	Multi-block Single Shot MultiBox Detector (SSD)	Novel Multi-block SSD architecture optimized for small objects in UAV remote sensing images.	Surveillance of railway scenes using Unmanned Aerial Vehicle (UAV), especially for small objects.
   Urban rail transit obstacle detection based on improved R-CNN	D. He, K. Li, R. Ma, Y. Qin, R. Ren, W. Yang and Z. Zou	Deep Learning/Im age Recognition method	Focuses on the complex and dynamic environment of urban rail transit.	Urban rail area and safe operation against specific obstacles.
   RailFOD23: A dataset for foreign object detection on railroad transmission lines	Z. Chen, J. Yang, Z. Feng and H. Zhu	N/A (Dataset/Eva luation)	Presents the RailFOD23 dataset (14,615 images) and uses large-scale models to synthesize data.	Foreign object detection on railroad power transmission lines(Catenary/Over head lines).
   Fast detection study of foreign object intrusion on railway track	H. Niu and T. Hou	Multi- background modeling + Multi- difference method	New detection method based on multi- background modeling and black/white pixel proportion to establish the dangerous area.	Improving detection accuracy and rapidity in real-time monitoring.
   Research on the method of foreign object detection for railway tracks based on deep learning	S. Ning, F. Ding and B. Chen	Deep Learning Method (General)	Aims to address inadequate real-time performance and diminished accuracy in detecting small objects on tracks.	Improving real-time performance and accuracy for small objects on railway tracks.

   Table 1: Comparison Table of Foreign Object detection in railway

   This table compares 20 papers on Railway Foreign Object Detection (FOD), contrasting early methods (Image Processing) with modern Deep Learning (DL). It highlights the use of YOLO-variants and attention mechanisms for real- time speed, alongside innovative techniques for small object detection, zero-shot anomaly detection, and tackling data scarcity.

4. RESEARCH CHALLENGES

   Despite various advancements in the foreign object detection in the railways using deep learning and computer vision techniques, several challenges still remain .One such challenge is the detection of partially occluded or small objects such as stones or debris on the railway tracks that are difficult to detect accurately. Another challenge is the lack of high computational resources that might affect the real time monitoring of the railway track and object detection. In addition, the availability of large and diverse railway datasets is limited, which affects the training and ability of the object detection models. Addressing these challenges is essential for developing effective railway systems.

5. CONCLUSION

Over the years, research in railway foreign object detection has evolved remarkably, moving from simple sensor-based systems to sophisticated AI-driven approaches. Early IoT and ultrasonic sensor methods, as explored by Aziz et al., offered low-cost, real-time monitoring but were limited by environmental challenges and short detection ranges. Classical image processing techniques improved visual recognition yet struggled under changing lighting conditions and complex track environments. The advent of deep learning models, including CNNs, Mask R-CNN, and YOLO variants, brought significant improvements in accuracy, small-object detection, and robustness, as shown by Gao, Meng, and Chen. Multi-stage frameworks and knowledge distillation enabled lightweight classifiers to work alongside powerful detectors, achieving a balance between speed and precision. Attention mechanisms, transformer layers, and BiFPN modules enhanced feature extraction, allowing reliable detection even under weak illumination, occlusion, or adverse weather conditions. UAV- based and camera-based monitoring solutions demonstrated practical real-time surveillance capabilities and highlighted the feasibility of edge deployment on devices like Jetson Nano. The creation of large datasets and synthetic augmentation techniques, such as those by Fan et al. and Chen et al., addressed data scarcity and improved model generalization across diverse scenarios. Semi-supervised and multimodal approaches further expanded the ability to detect anomalies without relying on extensive labeled data, making these systems more adaptable and scalable. Collectively, these studies illustrate a clear trend toward integrating visual AI, efficient architectures, and attention- enhanced models to develop fast, accurate, and reliable railway safety systems capable of preventing accidents and ensuring operational security.

I. REFERENCES

1. C. Tatimur, M. Karaköse, and E. Akn, Image processing based level crossing detection and foreign objects recognition approach in railways, International Journal of Applied Mathematics Electronics and

   Computers, Special Issue-1, pp. 1923, Sept. 2017.

2. L. Niu, Automatic identification method of foreign body intrusion in railway transportation track based on improved LeaderRank identification of key points, Egyptian Informatics Journal, vol. 30, no. 2, p. 100682, 2025.
3. Railway- CLIP: A multimodal model for abnormal object detection in high-speed railway, High-Speed Railway, vol. 3, no. 3, pp. 175260, Jul. 2025. [4 ] H. Niu, D. Feng, and T. Hou, Research on foreign object intrusion detection in railway tracks based on MSL-YOLO, Journal of Engineering and Applied Science, vol. 72, no. 1, p. 136, 2025.

1. W. Chen, S. Meng, and Y. Jiang, Foreign object detection in railway images based on an efficient two-stage convolutional neural network, Computational Intelligence and Neuroscience, vol. 2022.
2. S. Gao, A dense attention railway foreign object detection algorithm based on Mask R-CNN, IEEE Access, vol. 12, pp. 8576185772, 2024.
3. S. Ning, R. Guo, P. Guo, L. Xiong, and B. Chen, Research on foreign object intrusion detection for railway tracks utilizing risk assessment and YOLO detection, IEEE Access, vol. 12, pp. 175926175939, 2024.
4. X. Fan, Y. Zhang, and H. Liu, Track foreign object image augmentationusing PLCA-pix2pixGAN, Journal of Railway Science and Engineering, vol. 21, no. 5, pp. 987999, 2024.
5. T. Ye, J. Zhao, and L. Xu, Automatic railway traffic foreign object detection based on FFR-CNN, IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 2, pp. 1543

   1554, 2024.

6. Y. Yang, Z. Liu, J. Chen, H. Gao, and T. Wang, Railway foreign object intrusion detection using UAV images and YOLO-UAT, IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 5, pp. 6789 6801, 2021.
7. Z. Qi, D. Ma, J. Xu, A. Xiang, and H. Qu, Improved YOLOv5 based on attention mechanism and FasterNet for foreign object detection on railway and airway tracks, arXiv preprint arXiv:2403.08499, Mar. 2024.
8. X. Cai, YOLOv7-based research on foreign object intrusion detection on tracks, International Core Journal of Engineering, vol. 8, no. 12,

   pp. 117122, Dec. 2022.

9. T. Wang, Z. Zhang, F. Yang, and K.-L. Tsui, Intelligent railway foreign object detection: A semi-supervised convolutional autoencoder based method, arXiv preprint arXiv:2108.02421, Aug. 2021
10. Z. Chen, J. Yang, Z. Feng, and H. Zhu, RailSegVITNet: Lightweight VIT-based segmentation for railway foreign object detection, IEEE Access, vol. 12, pp. 123456123468, 2024.
11. Y. Li, H. Dong, H. Li, X. Zhang, B. Zhang, and Z. Xiao,

    Multi-block SSD based on small object detection for UAV railway scene surveillance, Chinese Journal of Aeronautics, vol. 33, no. 6, pp. 17471755, 2020.

12. Y. Li, H. Dong, H. Li, X. Zhang, B. Zhang, and Z. Xiao, Multi- block SSD based on small object detection for UAV railway scene surveillance, Chinese Journal of Aeronautics, vol. 33, no. 6, pp. 17471755, 2020
13. D. He, K. Li, R. Ma, Y. Qin, R. Ren, W. Yang, and Z. Zou, Urban rail transit obstacle detection based on improved R- CNN, Measurement, vol. 196, p. 111277, 202
14. Z. Chen, J. Yang, Z. Feng, and H. Zhu, RailFOD23: A dataset for foreign object detection on railroad transmission lines, Scientific Data, vol. 11, no. 72, pp. 112, 2024.