嵌設SURF演算法之粒子群聚最佳化法的多物體追蹤

Abstract

即時目標影像追蹤對於機器人視覺、監控系統等工業應用極為重要，因此本論文提出一種以粒子群聚最佳化演算法(Particle Swarm Optimization, PSO)為基礎之多目標物體追蹤法。作法上係以待追蹤物體之灰階直方圖做為物體的特徵，並產生多個粒子群體，再利用各個粒子所選取到之影像區間的灰階直方圖與目標物的灰階直方圖之差距作為適應值，利用PSO演算法具有記憶及全域搜尋和同步處理等優點，來搜尋各個目標物體。由於直方圖建模是由其目標物體影像大小來做建模，但是當目標物體因尺度(scale)變化時，其直方圖便不會跟著尺度變化而改變，所以會造成追蹤錯誤，因此本文將利用影像中之加速穩健特徵(Speeded Up Robust Features, SURF)，在PSO演算法中嵌入SURF演算法以提取目標物體影像並重新建模取得直方圖，使得PSO不會因為目標物體因尺度變化時得到錯誤的直方圖，而造成追蹤錯誤。

This paper presents a particle swarm optimization (PSO) based approach for multiple object tracking based on histogram matching. To start with, gray-level histograms are calculated to establish feature model for each of the target object. The difference between the gray-level histogram corresponding to each particle in the search space and the target object is used as the fitness value of the PSO algorithm. Multiple swarms are created, depending on the number of the target objects under tracking. Because of the efficiency and simplicity of the PSO algorithm for global optimization, target objects can be tracked as iterations continue. Experimental results confirm that the proposed PSO algorithm can rapidly converge, allowing real-time tracking of each target object. It is observed that tracking performance based on histogram modeling seriously deteriorates if the image size of the target objects varies in scale. To solve this problem, Speeded Up Robust Features (SURF) is embedded into the tracking algorithm extract the accurate position of the target objects in the image, where histogram models of the target objects are re-generated for tracking during the evolution of the PSO algorithm. As a result, tracking performance is significantly improves, as demonstrated in the experimental results.