Search

The 3D perception system functions as one of the core components for the safe operation of autonomous-rail rapid tram. Addressing the limitations of LiDAR systems used for autonomous-rail rapid tram in complex urban environments, such as challenges in perceiving distant objects, insensitivity to color, and potential failure, this paper proposes a pure-vision monocular 3D perception system for autonomous-rail rapid tram. Comprising data preprocessing, model training and model deployment, this vision-based scheme with a full coverage from data collection to on-board deployment enables the 3D perception of obstacles around autonomous-rail rapid tram, thereby improving the operational reliability in complex urban environments. The test results based on the autonomous-rail rapid transit dataset and Waymo open dataset show that the system can effectively perceive the complex road scenes in autonomous-rail rapid transit, achieving a final 3D average precision (AP) of 0.53, reasoning a single-frame of image about 56 ms, and meeting the real-time requirements of autonomous-rail rapid tram for the obstacle perception algorithm.

Aiming at the problems of poor detection effect and low tracking stability of multi-type vehicles in complex mining environment due to the similarity of operating vehicles and background images, this paper proposes a multi-category and multi-target real-time detection and tracking algorithm for operating vehicles in complex mining environments. The model framework is constructed based on the lightweight backbone network YOLO combined with the multi-scale feature fusion module. The model uses DIoU as loss function, uses K-means clustering to regress the size of candidate frame, and learns image features through the lightweight backbone network. On this basis, the features of multi-category work vehicle targets are used as similarity metric, combined with Mahalanobis distance metric and cosine metric that characterize motion information for cascading matching, and IoU matching and Kalman filtering are connected in series to confirm the trajectory and the real-time tracking of multiple work vehicles. Experimental results show that the average vehicle detection accuracy of the algorithm mAP@ 0.5-0.95 is 58.40%, the multi-target tracking accuracy reaches 82.60%, and the image processing time per frame is 26.5 ms, which can effectively perform real-time detection and tracking of working vehicles.

Track component defect detection has long faced the problem of scarcity of defect samples. In the case of insufficient defect samples, the existing deep learning methods are prone to overfitting the model, and the generalization performance is poor, making it difficult to meet the actual track inspection requirements. To this end, this paper proposes a cascaded track component defect detection method which combines instance segmentation, fine-grained image classification and traditional image processing. Firstly, an improved fast instance segmentation network is proposed to locate fasteners and defective rails. Secondly, in view of the scarcity of defective fastener samples, a method of mask analysis and quantitative comparison of fastener contour lengths is proposed to realize the low-sample condition. At the same time, in view of the problem that the categories of rail defects are complex and subtle and easy to mix, a two-stage defect detection network of "instance segmentation + fine-grained image classification" is proposed, and label smoothing technology is introduced to improve model adaptability and classification accuracy. Experimental results show that, in the case of scarce samples, the defect detection method proposed in this paper has an accuracy rate of 95.7% for fastener breakage and loss detection, and 98% and 95% for rail defect detection rate and defect sub-classification, respectively. It meets the demand for track component defect detection.