基於Kinect之運用粒子濾波器於多人追蹤與交疊處理

Abstract

本篇論文提出了一套利用粒子濾波器為核心之人物偵測與追蹤演算法並基於Kinect感測器的完整系統，其中利用前景人物與深度特徵之基礎整合，偵測出人形候選物件，並根據人體寬度與深度之位置關係進行人物判定。 在論文貢獻上，我們設計了一套即時系統配合粒子濾波器演算法進行人物追蹤，並搭配各項人物特徵，包含位置、色彩、深度等資訊執行追蹤，除此之外，不同於其他多數的人形追蹤研究，本論文針對遮蔽與人物交疊等特殊情況，有更進一步的研究，並根據適應不同狀況的方法，搭配人物追蹤與其深度資訊加以進行狀況處理與問題解決，另外，針對追蹤中人物的姿勢變換與深度位置的改變，處理追蹤目標之邊界，使系統能完全追蹤人物目標之位置與影像中對應之身體寬度及高度等各項資訊。 實驗結果中顯示不論是在場景光線變化或完全不足的情況下，皆能利用Kinect支援之深度資訊進行演算法的輔助，以克服光線造成之環境變化與挑戰，並提供完整無誤的追蹤結果。

This paper presents a novel human detection and tracking system with the particle filter approach based on sensors of Kinect. The human detection module extracts the human position by integrating foreground extraction and the depth image. Moreover, the human bodies are judged according to the scaled features with depth information. The advantage of this paper is we consider a real-time particle filtering approach to track multi-human with parameters in several features including the position, color and depth information. Furthermore, different from the most tracking system, we present the occlusion handling approach by the cooperation of the depth and tracking information, and the scaling processes with the different human postures and the positions of depth. Experimental results show that the proposed human tracking system achieves good results in the challenged video. Even in the all dark scenes and light changing situations, the algorithm we presented can still work in a good performance.