Occluded Pedestrian Detection Based on Joint Attention Mechanism of Channel-wise and Spatial Information

Abstract

Pedestrian detector performance is damaged because occlusion often leads to missed detection. In order to improve the detector's ability to detect pedestrian, a single-stage detector based on feature-guided attention mechanism is proposed. Firstly, a feature attention module is designed, which preserves the association between the feature channels while retaining spatial information, and guides the model to focus on visible region. Secondly, the attention module is used to fuse shallow and deep features, then high-level semantic features of pedestrians are extracted. Finally, pedestrian detection is treated as a high-level semantic feature detection problem. Pedestrian location and scale are obtained through heat map prediction, then the final prediction bounding box is generated. This way, the proposed method avoids the extra parameter settings of the traditional anchor-based method. Experiments show that the proposed method is superior to other comparison algorithms for the accuracy of occlusion target detection on CityPersons and Caltech pedestrian database. At the same time, the proposed method achieves a faster detection speed and a better balance between detection accuracy and speed.