A Multi-Pedestrian Tracking Method Based on Iterative Filtering and Observation Discrimination

Abstract

The invention discloses a multi-pedestrian tracking method based on iterative filtering and observation and discrimination. The video image to be detected is subjected to three iterations of component combination detection to reduce the missed detection rate; the grayscale of the newly added target head retained after the third iteration is calculated. The histogram area overlap ratio between the histogram and the average grayscale histogram of the head image block after the second iteration, filtering out the target head detection frame whose overlap ratio is less than the set threshold, effectively suppressing the effect of false detection frames on the detection performance Retain the influence of reliable target detection frame, which is beneficial to improve the detection accuracy; further extract the local observable area of ​​the mutual occlusion or incomplete detection target, obtain the center and scale information of the multi-target observable area, and establish an observation data set; according to Observation data set and target trajectory confidence to achieve tracking. The invention can be applied to the fields of artificial intelligence, intelligent robot and intelligent video surveillance.

Description

technical field [0001] The invention relates to a multi-pedestrian tracking method based on iterative filtering and observation discrimination, belongs to the field of computer vision, and is mainly used in artificial intelligence, intelligent robots and intelligent video surveillance. Background technique [0002] Multi-target tracking is one of the research hotspots in the field of computer vision and intelligent video information processing. It has a wide range of applications in public security monitoring and management, medical image analysis, behavior understanding, and visual navigation. At present, the main concerns of scholars at home and abroad focus on the improvement of tracking robustness and accuracy in complex scenes such as interference of similar features between targets, blurred appearance and occlusion. [0003] FelzenszwalbPF et al. first proposed the Deformable Parts Model DPM (Deformable Parts Model) in the article "Object Detection with Discriminativel...

AI Technical Summary

Problems solved by technology

This method has a small amount of calculation and meets the real-time tracking requirements. However, after multiple targets are occluded and separated, there will be problems of target identity confusion and trajectory misjudgment.

In the article "Robust Online Multi-object Tracking Based on Tracklet Confidence and Online Discriminative Appearance Learning" (TC-ODAL) published in IEEE Conference on Computer Vision and Pattern Recognition (2014:1218-1225), Bae S H et al. The quantitative linear discriminant appearance algorithm solves the problem of identity confusion in the process of multi-target tracking and improves the performance of multi-target tracking. However, when the target is lost and reappears, the algorithm has the problem of reinitializing the new identity label.

In the article "High-Speed ​​Tracking-by-Detection Without Using Image Information" published by International Workshop on Traffic and Street Surveillance for Safety and Security at IEEE Avss (2017:1-6), Bochinski E et al. The change proposes a multi-target tracking algorithm based on the target region of interest. The algorithm can run at a rate of 100,000 frames per second, but in the case of missing detection, it will reduce the tracking performance

Images