Enhanced Road Vehicle Object Detection Based on Improved Deformable DETR
Hua Yin, Liang Chen
Abstract
An improved Deformable DETR model is proposed to address the problem of insufficient accuracy, complex structure, and excessive computational burden in traditional road vehicle object detection. Firstly, single-scale features are separately processed through an encoding layer and then concatenated, effectively reducing the parameters and computation of the model. Secondly, a channel attention mechanism is proposed to enhance the encoder, strengthening its ability to aggregate feature information and improving the model's accuracy and robustness. Finally, the query selection mechanism is improved, transitioning from the original position and content queries to enhancing only the position query while maintaining the learnability of the content query. Experiments conducted on the BIT-Vehicle dataset demonstrate that the proposed model based on improved Deformable DETR outperforms the original model, with a 3.7<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> increase in mAP, a 5% reduction in parameters, and a 24.5% decrease in computational loads.