Gradient-Guided Joint Representation Loss With Adaptive Neck for Train Crash Detection

Conducting real-world train crash experiments is the most straightforward and effective method to research the train’s crashworthiness and enhance passive safety protection. As a non-contact measurement method, high-speed camera can efficiently capture the evolving motion patterns of trains under the high-speed states. Traditional data extraction methods rely on expert-based manual annotations, which are susceptible to factors such as illumination changes, scale variance, and shock debris. Inspired by the tremendous success of Deep Neural Networks in the computer vision community, we firstly collect 75 real-world train crash scenes and manually annotated them to form the Crash2024 dataset, enriching the community’s data resources. Moreover, we propose a novel Gradient-guided Joint representation loss with Adaptive neck Detection network (GJADet). At the macro level, we embed the adaptive module into the Path Aggregation Feature Pyramid Network, which combines multiple self-attention mechanisms to achieve scale-awareness, spatial-awareness, and task-awareness, improving the detector’s representation ability and alleviating the dense-small characteristics of the ‘point’ class without significant computational overhead. At the micro level, due to the extreme imbalance of ‘point’ class compared to other classes, we propose a gradient-guided joint representation classification loss to mitigate the long-tailed detection issue. Moreover, the classification and regression are joint representation to maintain consistency between training and inference phase. On the Crash2024, GJADet achieves the performance improvement of 3.5AP and significantly alleviate the accuracy loss problem for rare categories. Our code are open source at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/YanJieWen/GJADet-crash-pytorch</uri>.