A Lightweight Real-time Transformer for Pedestrian Detection

In current pedestrian detection algorithms, the occlusion between pedestrians in dense scenes causes such problems as missing detection, false detection, and high cost of dense information extraction. Therefore, a lightweight real-time Transformer-based pedestrian detection model, called LRTM, is proposed. Initially, the re-parameterized FasterNet is utilized as the fundamental unit to extract preliminary feature information from the backbone network. Subsequently, a Deformable Inter-Scale Feature Interaction (DIFI) module is introduced, incorporating channel shuffling and grouped convolution as the feature extractor within the encoder. Furthermore, a spatial decay matrix based on Chebyshev distance is proposed to enhance the fusion of contextual features. Finally, the proposed model LRTM is trained and validated using the CrowdHuman dataset. Experimental results indicate that, compared to the baseline model RT-DETR, a reduction of 21.9% in parameter count is achieved, along with an increase of 2.7 percentage points in mAP0.5 and an improvement of 3 percentage points in mAP 0.5:0.95.