Redundant contextual feature suppression for pedestrian detection in dense scenes

Abstract

Pedestrian detection is one of the important branches of object detection, with a wide range of applications in autonomous driving, intelligent video surveillance, and passenger flow statistics. However, these scenes exhibit high pedestrian density, severe occlusion, and complex redundant contextual information, leading to issues such as low detection accuracy and a high number of false positives in current general object detectors when applied in dense pedestrian scenes. In this paper, we propose an improved Context Suppressed R-CNN method for pedestrian detection in dense scenes, based on the Sparse R-CNN. Firstly, to further enhance the network’s ability to extract deep features in dense scenes, we introduce the CoT-FPN backbone by combining the FPN network with the Contextual Transformer Block. This block replaces the $3\times 3$ convolution in the ResNet backbone. Secondly, addressing the issue that redundant contextual features of instance objects can mislead the localization and recognition of object detection tasks in dense scenes, we propose a Redundant Contextual Feature Suppression Module (RCFSM). This module, based on the convolutional block attention mechanism, aims to suppress redundant contextual information in instance features, thereby improving the network’s detection performance in dense scenes. The test results on the CrowdHuman dataset show that, compared with the Sparse R-CNN algorithm, the proposed algorithm improves the Average Precision (AP) by 1.1% and the Jaccard index by 1.2%, while also reducing the number of model parameters. Code is available at https://github.com/davidsmithwj/CS-CS-RCNN.