A multi-view pedestrian detection method based on multi-view Bayesian network

Abstract

The present invention proposes a pedestrian detection method and device based on a perspective Bayesian network model, which can detect and locate pedestrians in densely populated scenes monitored by multiple cameras. It is not only suitable for scenes with good foreground extraction, but also for scenes with poor foreground extraction but can be detected by pedestrian detectors. The method of the invention includes a single-view processing step, a base plane mapping step, a multi-view fusion step, an inverse mapping step and a final detection result output step.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and relates to a monitoring video pedestrian detection and analysis method, in particular to a pedestrian detection method based on a multi-view Bayesian Network (MvBN) model, and a device for realizing the method. Background technique [0002] As surveillance cameras are widely used in all aspects of people's lives, surveillance video data has shown explosive growth. Pedestrian detection is one of the most important steps in other surveillance video analysis, which provides the basis for other surveillance video analysis applications, such as object tracking, event monitoring, etc. The detection accuracy directly affects other surveillance video analysis applications. Compared with ordinary single-view pedestrian detection, multi-view pedestrian detection can more effectively deal with false detection and missed detection caused by occlusion. And multi-view pedestrian detection can not...

AI Technical Summary

Problems solved by technology

This method needs to find the main axis in the Unicom area of ​​the single-view foreground, which has relatively high requirements for the foreground and is not robust enough.

Moreover, this method only uses foreground information, but does not use texture features

Therefore, in scenes with inaccurate foreground feature extraction (such as strong lighting changes, etc.), the processing effect of this method will be greatly reduced.

[0006] The foreign literature "Multi-camera Pedestrian Detection with a Multi-view BayesianNetwork Model" (published in BMVC 2012, Proceedings of the British Machine Vision Conference, Peixi Peng, Yonghong Tian, ​​Yaowei Wang, Tiejun Huang, 2012) proposed a multi- As for the pedestrian detection method, only single-view foreground detection can be used in this document, and the detection results of single-view foreground detection and the detection results of pedestrian detectors cannot be used in single-view detection.

Images