Exploiting key points supervision and grouped feature fusion for multiview pedestrian detection. (November 2022)

• Record Type: Journal Article
• Authors: Gao, Xin; Xiong, Yijin; Zhang, Guoying; Deng, Hui; Kou, Kangkang
• Is Part Of: Pattern recognition. Volume 131(2022)
• Publication Date: 2022-11
• Subjects: Multiview aggregation -- Pedestrian detection -- Key points -- Grouped feature fusion; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108866
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We introduce a method to accomplish multiview pedestrian detection using key points regression and the correlation of overlapping fields of multiple views, which achieves effective multiview features aggregation under the action of three types of key points and grouped feature fusion modules. The proposed key points supervision method regresses pedestrians into three points, which can effectively handle pedestrian detection under occlusion. The proposed grouping feature fusion module is the first attempt to apply the correlation of multiview overlapping fields to multiview feature aggregation, the feature enhancement and fusion for this region can help reduce object ambiguity. Compared to state-of-the-art methods, our method achieves superior performance with only a small increase in computation. Abstract: Multiview pedestrian detection detects pedestrians based on the perception of the same environment from multiple perspectives. This task requires feature extraction in a single view with occlusion and aggregation of multiview information. However, existing research is limited by the local occlusion and the multiview feature stitching method, which cannot perform multiview aggregation efficiently. This paper introduces a network that utilizes key points supervision and grouped feature fusion to address these challenges. It uses key points to regress pedestrians in a single view, and augments the pedestrian consistency information in overlapping views by a grouped feature fusion module. Specifically, the proposed key points supervision effectively alleviates false negatives due to occlusion, and the grouped feature fusion module enhances pedestrian location features by computing the similarity and spatial correlation of overlapping views after single view projection to the ground plane, thereby reducing target ambiguity. Quantitative and qualitative results show that the proposed method can reduce false negatives and false positives in multiview pedestrian detection and achieve efficient multiview feature aggregation. Compared to state-of-the-art methods, the proposed model achieves superior performance, achieving the highest MODA of 92.4 and 93.9 on Wildtrack and MultiviewX datasets, respectively. We believe, to the best of our knowledge, that this approach offers a new optimization idea for multiview aggregation.