基於粒子群聚演算法的多行人追蹤

Abstract

行人的偵測與追蹤是近年來相當重要的研究項目。一般常見的行人追蹤演算法大多是利用大量的運算時間來取得高準確率，或者是犧牲準確率而得到快速追蹤的結果。直至目前為止，尚未有快速且準確的演算法來偵測追蹤移動中的行人。因此，本論文採用運算速度更為快速的粒子群聚(Particle Swarm Optimization, PSO)演算法搭配色彩直方圖 (Color Histogram) 做為擷取影像特徵方法，以達到速度與準確率的最佳平衡。本研究的方法共分成四個階段：首先利用Histogram of Oriented Gradient (HOG) 針對輸入影像進行「行人偵測」找出行人位置，其次對該影像進行「影像前處理」來降低光影的影響，接著計算影像的色彩直方圖進行特徵距離相似度計算，最後套用至PSO演算法的適應函數進行行人追蹤。本研究的實驗從七大影像資料庫中選取不同的移動方式的影像 (如左右橫向移動、畫面上隨機移動及深度的移動等) 來驗證PSO演算法的速度與準確率。結果顯示，在花費較少時間計算的情況之下 (0.0784 ~ 0.0906 s) ，PSO演算法可達到與其他演算法一樣甚至較高的準確率 (Multiple Object Tracking Accuracy，MOTA：69.88 ~ 86.54%；Multiple Object Tracking Precision，MOTP：77.43 ~ 84.76%) 。若影像資料沒有受到部分或全部遮蔽的干擾，PSO演算法的追蹤準確率可維持在80% 以上。研究結果證實與現有的演算法相比，PSO演算法不僅大幅縮短了演算時間，更展現優異的追蹤準確率，顯示此演算法可達到速度與準確率的最佳平衡。未來研究將繼續改良現有的PSO演算法，以提升分析部分或全部遮蔽影像的準確率。Human detection and tracking has been an important research topic in recent years. Most commonly used methods of human tracking either require a large amount of computing time to reach high accuracy, or sacrifice the accuracy to obtain rapid tracking results. To date, there is no algorithm which can fast and accurately detect and track pedestrians. In light of this, the current study employed the Particle Swarm Optimization (PSO) algorithm with Color Histogram methods, aimed to develop an optimized algorithm for human detection and tracking. The proposed algorithm include four steps: 1) Histogram of Oriented Gradient (HOG) was used to detect the locations of pedestrians; 2) Image Pre-processing was used to remove the influence of light and shadow; 3) Color Histogram of the image was computed to calculate the similarity of feature points; and 4) Fitness Function of PSO algorithm was computed to proceed Human tracking. To examine the speed and accuracy of the proposed algorithm, images with pedestrians in different moving ways (Lateral movement, randomize movement and depth movement) were collected from seven image databases. The findings suggest that compare to other algorithms, the PSO algorithm can reach similar or better accuracy (Multiple Object Tracking Accuracy, MOTA：69.88 ~ 86.54%；Multiple Object Tracking Precision, MOTP：77.43 ~ 84.76%) in shorter computing time (0.0784 ~ 0.0906s). Even if the image data were interfered by partial or entire occlusion, the tracking accuracy of the PSO algorithm can still reach 80% or above. The findings reveal that compared with other existed algorithms, the PSO algorithm not only significantly reduces the computing time, but also can reach excellent tracking accuracy. Future studies are warranted to improve the PSO algorithm tacking accuracy when process image data with partial and entire occlusion.