Full occlusion pedestrian tracking method and device based on hybrid system

Abstract

The present invention provides a fully-shielded pedestrian tracking method based on a hybrid system. According to the pedestrian movement features and pedestrian features required learning in the tracking, a corresponding tracking model and an identification model are constructed, when pedestrians are not shielded, the tracking model learns pedestrian features while tracking the pedestrians, the shielded pedestrians are sensed through the sensing condition of the model transition result and the tracking model is switched to the identification model when the pedestrians are shielded; the identification model utilizes the learned pedestrian features to continuously identify pedestrians with tracking loss from the detection result according to the identification condition of the model transition rule; and when the identification model identifies the pedestrians with tracking loss, and the states of tracking pedestrians are reset through the reset condition of the model transition rule and the is identification model switched to the tracking model for going on tracking so as to realize pedestrian tracking in the full shielding condition. The design is reasonable, the model driving and the data driving algorithm employ the idea of the hybrid system to perform integration so as to solve the actual problem so as to greatly improve the tracking accuracy and success rate.

Description

technical field [0001] The invention relates to a technology in the field of computer vision monitoring target tracking, in particular to a hybrid system-based full-occlusion pedestrian tracking method and device thereof. Background technique [0002] Traditional single pedestrian tracking mainly has two categories of methods: target representation and localization, filtering and data association. The commonly used pedestrian tracking algorithm is the first type of method. These methods can be divided into three sub-categories: generative tracking, discriminative tracking, and hybrid detection and tracking. The generative algorithm learns a representative model to represent the tracking target and uses the model to search for the image area that minimizes the reconstruction error; the discriminative algorithm regards the tracking problem as a binary classification problem in the local search area, and uses the learned classifier in the current Frame search finds the most d...

AI Technical Summary

Problems solved by technology

However, this technology 1) due to the defects of the pedestrian detection algorithm of HOG-LBP descriptor plus SVM classifier, it is often necessary to add training samples for tracking pedestrians when selecting positive and negative samples, which is not conducive to the detection of unknown pedestrians; 2) in solving occlusion The problem did not consider the situation that pedestrians fall outside the particle filter prediction area; 3) The algorithm does not have a track termination mechanism, and the template is still updated after the pedestrian is occluded, which will inevitably lead to template update errors and subsequent pedestrian tracking cannot be performed; 4) According to Algorithm description If a pedestrian appears in a position not predicted by the particle filter after being fully occluded for a period of time, it is easy to be identified as a new target

Images