A dual-channel SSD pedestrian detection algorithm based on feature fusion

In order to solve the phenomenon of false detection and missed detection in the pedestrian counting algorithm caused by the change of occlusion and illumination in crowded scenes, this paper proposes a dual-channel SSD pedestrian detection algorithm based on feature fusion, which improves pedestrian detection under illumination and occlusion conditions. In the dual-channel SSD network structure, the Conv4_3 and FC7 layers in the color image channel are fused with the Conv4_3 and FC7 layers in the depth image channel in Concat or Eltwise ways to obtain the optimal dual-channel SSD network model. Then, the Conv4_3_Fuse layer in the network is fused with the Conv10_2_Fuse and Conv11_2_Fuse layers to fully learn the feature information of pedestrian heads. The experimental results show that the improved algorithm is tested on the TVHeads data set, and the detection accuracy obtained is 95.4%, which is 13.16% higher than that of the SSD algorithm, which improves the problem of missed detection caused by illumination changes and occlusion, and enhances the detection of pedestrians head recognition.