With the rapid development of high-speed railways, any objects intruding railway clearance will do great threat to railway operations. Accurate and effective intrusion detection is very important. An original Single Shot multibox Detector (SSD) can be used to detect intruding objects except small ones. In this paper, high-level features are deconvolved to low-level and fused with original low-level features to enhance their semantic information. By this way, the mean average precision (mAP) of the improved SSD algorithm is increased. In order to decrease the parameters of the improved SSD network, the L1 norm of convolution kernel is used to prune the network. Under this criterion, both the model size and calculation load are greatly reduced within the permitted precision loss. Experiments show that the mAP of our method on PASCAL VOC public dataset and our railway datasets have increased by 2.52% and 4.74% respectively, when compared to the original SSD. With our method, the elapsed time of each frame is only 31 ms on GeForce GTX1060.

Video surveillance-based intrusion detection has been widely used in modern railway systems. Objects inside the alarm region, or the track area, can be detected by image processing algorithms. With the increasing number of surveillance cameras, manual labeling of alarm regions for each camera has become time-consuming and is sometimes not feasible at all, especially for pan-tilt-zoom (PTZ) cameras which may change their monitoring area at any time. To automatically label the track area for all cameras, video surveillance system requires an accurate track segmentation algorithm with small memory footprint and short inference delay. In this paper, we propose an adaptive segmentation algorithm to delineate the boundary of the track area with very light computation burden. The proposed algorithm includes three steps. Firstly, the image is segmented into fragmented regions. To reduce the redundant calculation in the evaluation of the boundary weight for generating the fragmented regions, an optimal set of Gaussian kernels with adaptive directions for each specific scene is calculated using Hough transformation. Secondly, the fragmented regions are combined into local areas by using a new clustering rule, based on the region’s boundary weight and size. Finally, a classification network is used to recognize the track area among all local areas. To achieve a fast and accurate classification, a simplified CNN network is designed by using pre-trained convolution kernels and a loss function that can enhance the diversity of the feature maps. Experimental results show that the proposed method finds an effective balance between the segmentation precision, calculation time, and hardware cost of the system.

Foreign object intrusion is a great threat to high-speed railway safety operations. Accurate foreign object intrusion detection is particularly important. As a result of the lack of intruding foreign object samples during the operational period, artificially generated ones will greatly benefit the development of the detection methods. In this paper, we propose a novel method to generate railway intruding object images based on an improved conditional deep convolutional generative adversarial network (C-DCGAN). It consists of a generator and multi-scale discriminators. Loss function is also improved so as to generate samples with a high quality and authenticity. The generator is extracted in order to generate foreign object images from input semantic labels. We synthesize the generated objects to the railway scene. To make the generated objects more similar to real objects, on scale in different positions of a railway scene, a scale estimation algorithm based on the gauge constant is proposed. The experimental results on the railway intruding object dataset show that the proposed C-DCGAN model outperforms several state-of-the-art methods and achieves a higher quality (the pixel-wise accuracy, mean intersection-over-union (mIoU), and mean average precision (mAP) are 80.46%, 0.65, and 0.69, respectively) and diversity (the Fréchet-Inception Distance (FID) score is 26.87) of generated samples. The mIoU of the real-generated pedestrian pairs reaches 0.85, and indicates a higher scale of accuracy for the generated intruding objects in the railway scene.

Railway scene status perception is the foundation for ensuring the efficient and safe operation of trains. Achieving comprehensive railway scene perception requires not only detecting objects but also understanding the relationships between them. However, current railway scene perception technologies struggle with the latter, while general object and relationship perception methods perform poorly when facing railway scene with significant structural features and small object characteristics. Thus, this paper proposes Rail-former, the first status perception framework specifically designed for railway scene. Rail-former operates based on two key components: the Railway Feature Enhancement Module (RFEM) and the Cross-Coupled Transformer (CCTF). RFEM first leverages a cluster-based normalization evaluation method to condense railway scene structural features in spatial domain. It then applies high-dimensional information compensation to enhance small-object feature representation. The CCTF implements a cross-coupled structure between object decoder and triplet decoder based on a novel interactive attention mechanism, rather than conventional single-decoder or parallel dual-decoder structures, enabling guided enhancement of both object features and relationship features in railway scene. Additionally, we introduce a multimodal matching loss function that incorporates masks, categories, and bounding boxes to mitigate overfitting to a single modality during training. To the best of our knowledge, Rail-former is the first work in the railway domain that leverages panoptic parsing results to reveal relationships. Extensive experiments conducted on both our self-constructed railway dataset and a public dataset demonstrate the superior performance of our approach, achieving accuracy rates of 49.9% and 37.4%, surpassing the best competing methods by 7.8% and 1.1%, respectively.