Semantic enhanced guide feature reconstruction for occluded pedestrian detection

In pedestrian detection, the inability to extract effective features due to pedestrians being severely occluded is one of the main reasons for missing pedestrian detection. To solve this problem, a semantic enhanced guided feature reconstruction algorithm for occlusion pedestrian detection is proposed. Firstly, the semantic feature enhancement module is designed based on the dependency between space and channel, and the global context information is established to enhance the feature of occlusion of pedestrians. Secondly, in order to focus on the visible area of pedestrians, the adaptive feature reconstruction module is used to generate the semantic segmentation map, and adaptively adjust the effective weight of the channel, enhance the distinguishability of pedestrians and background. Finally, the multi-level feature map is obtained by multi-level cascade two modules of semantic feature enhancement and adaptive feature reconstruction, and the multiple features are fusion for the final pedestrian analysis detection. On the challenging pedestrian detection benchmark CityPersons and Caltech, experimental results show that the proposed method achieves the missed rate of 47.28% and 44.04%, respectively, which effectively robust compared with other methods in the detection of occluded pedestrian.