LW-DETR: a lightweight transformer-based object detection algorithm for efficient railway crossing surveillance

Abstract Object detection at coal transportation railway crossings is crucial for accident prevention and traffic efficiency improvement. However, the application of existing methods on resource-constrained devices has seldom been considered. To address these challenges, in this paper, we propose a lightweight railway crossing object detection algorithm based on the Transformer framework, referred to as Light-Weight DEtection TRansformer (LW-DETR). In this algorithm, the Paddle Paddle-Lightweight CPU Convolutional Network (PP-LCNet) is employed as the backbone network, where standard convolution is combined with depthwise separable convolution for multi-scale feature extraction. Furthermore, the cross-scale feature fusion module is optimized to reduce redundant calculations and enhance feature fusion efficiency. Moreover, the Scylla-Intersection over Union loss function is introduced to comprehensively evaluate bounding box similarity, thereby improving object detection accuracy. Ablation experiments conducted on a modified Pascal Visual Object Classes (Pascal VOC) dataset demonstrate that LW-DETR, while maintaining acceptable detection accuracy, achieves a 135.3% increase in frames per second, a 71.7% reduction in parameters, and a 73.7% decrease in computational load, leading to effective lightweight performance. Comparative experiments with other popular object detection algorithms further confirm that LW-DETR significantly enhances detection speed while maintaining high accuracy, considerably reducing model size and validating the effectiveness of these improvements.