A Multiview Approach to Tracking People in Crowded Scenes Using Fusion Feature Correlation

Kai Chen, Yujie Huang, Ziyuan Wang

Abstract: Most of the current tracking methods for multi-target pedestrian tracking are unable to solve the problem where the tracking targets are blocked and reappears after disappearing in the camera perspectives, which brings great challenges to its practical application. To tackle this problem in dense crowds, we propose a multi-target pedestrian tracking method based on fusion feature correlation under multi-vision: Updating the pedestrian feature pool based on GMM to reduce the feature pollution; Then dynamically calculating the similarity threshold of target features based on K-means algorithm; Use the idea of voting to match pedestrian features and determine the addition and reappearance of pedestrians. The results on open dataset Shelf show that our method improve the accuracy and success rate of tracking under the condition of occlusion and reappearance after disappearance.