Pedestrian Detection Based on Lightweight Improvement YOLOX

Aiming at the problems of large amount of parameters, high computation of floating point calculations and large model volume when YOLOX model is applied to pedestrian detection, a lightweight improvement based on YOLOX is proposed. Firstly, the SA-Ghost2 module is proposed to replace the module used for feature extraction in the backbone network, compressed model volume and reduced computation. Secondly, add a CBR operator fusion block at the bottom of dark layer, the expressive ability and overall performance of the model is improved without adding extra computation. Finally, the original Loss function is improved to Focal Loss to accelerate model convergence and improve accuracy. The experimental results show that, compared with the original model, the amount of parameters, floating point operations and the model volume of the improved model are reduced by 32%, 17.4% and 31.4% respectively, and the experimental effect of mAP _ 0.5 on CityPersons dataset is increased by 1.4 percentage points.