A Lightweight Pedestrian Detection Model Integrating Attention Mechanism

Aiming at the detection requirements of medium and small size objects and occluded objects in pedestrian detection tasks, a lightweight YOLOv5 pedestrian detection algorithm fused with attention mechanism is proposed. Firstly, a DSP (DSC-Stage-Partial) module based on depthwise separable convolution is designed, which shows higher accuracy in the detection of small targets and occlusions while reducing the number of model parameters. Secondly, the SE attention mechanism is added in front of the SPPF layer of the backbone, which can better improve the feature representation ability, so that the network can locate and identify the targets more accurately. Then, a more lightweight GhostC3 module is used in the Neck part, which greatly reduces the network parameters and calculation while ensuring the accuracy of detection. Finally, Alpha IoU is used as the loss function to make the model better adapt to objects of different scales, thereby improving the robustness of the model. On VOC and MIT datasets, the images of pedestrians are selected for training. The detection accuracy reaches 86.7<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>, which is 0.9% higher than original model in mAP0.5 and 2.7% higher than original model in mAP0.5:0.95, compared with the original model, the number of parameters is reduced by 1.3%, and the calculation amount is reduced by 15.19%. At the same time, the comparison experiment of the proposed method with the existing mainstream target detection methods shows that it has higher detection rate for pedestrian targets, lighter model and faster speed, which has certain practical application value.