Pedestrian Image Occlusion Detection Method Based on Recurrent Adversarial Generative Network

Abstract

The invention discloses a pedestrian image occlusion detection method based on a cyclic confrontation generation network. The steps are: (1) randomly adding occlusion image blocks to an unoccluded pedestrian image to generate an occlusion pedestrian image and corresponding pedestrians with occlusion marks image; (2) take the above-mentioned pedestrian image with occlusion and the corresponding pedestrian image with occlusion label as two data domains, train the cyclic confrontation generation network model, and construct the mapping between the occlusion image and the occlusion label; (3) for the current For the input image to be labeled, use the trained cyclic confrontation generation network model to generate the occlusion label corresponding to the image, and at the same time segment the image to be labeled to obtain several superpixel blocks. According to the occlusion label information in each superpixel block, The region growing method is used to get the final occluded region. The method of the invention can model different types of occlusions in the absence of occlusion labels, and obtain pixel-level occlusion area detection results.

Description

technical field [0001] The invention relates to an occlusion detection method in the field of image analysis, in particular to a method for detecting an occlusion area in a pedestrian image by using a cyclic confrontation generation network. Background technique [0002] Pedestrian re-identification aims to match the same pedestrian target from images taken from different perspectives and at different times. This problem is a very important problem in video security monitoring, especially for the tracking and positioning of key people, which has very practical significance. However, in the mainstream pedestrian re-identification task, the matching images are all pedestrian images acquired by pedestrian detectors, and these detected images are often accompanied by occlusions. The occlusion in the image will greatly interfere with the original pedestrian texture features, which greatly reduces the accuracy of the subsequent pedestrian re-identification task. In order to avoi...

AI Technical Summary

Problems solved by technology

Since the mainstream general-purpose object detection method extracts a rectangular frame, it is slightly different from pixel-level tasks such as occlusion detection. In practical applications, many post-processing steps are inevitably involved, and the robustness of the model cannot be guaranteed. sex

Images