Research on Multiscale Pedestrian Detection Algorithm
Zimiao Song, Hongmei Jin, Zhanli Li · 2022 7th International Conference on Image, Vision and Computing (ICIVC) · 2022
To solve object detection issues in the field of surveillance video, including a low recognition rate and a high false alarm rate caused by the great change of scale and different degrees of occlusion, we propose a multiscale pedestrian detection method. A Res-spp module based on residual structure is proposed to fuse the local and global features of objects after the backbone network. Besides, the CIOU_LOSS is employed as an object box position loss to promote the stability of the object’s bounding boxes regression. In addition, we make a neck structure that involves Weight_Conat mechanism designed for cross-layer multiscale feature fusion connection, which is based on learnable weights. It aims to strengthen the capability of network multiscale feature integration. Experimental results on VOC pedestrian dataset and widerperson dataset not only show the accuracy of the proposed method, 87.9% and 86.69% respectively, but also demonstrates a noticeable effect on the improvement of transformation of different scale and the occlusion problem of pedestrians.