Abstract: A key-frame-based bidirectional tracking algorithm is proposed to track a large number of high-density cell population. Each cell individual is marked manually in middle frame. The cell texture features are used to track cells across frames with fusing cell kinematics information and morphological information, which can resolve frequent cell overlap and occlusion problems. Trajectory fragments are inevitable because of over-congestion. The first and last frame is then selected as key frame. Re-tracking the missed cells can effective link these trajectory fragments. GPU is used to accelerate the system performance. Experimental results show that this algorithm can track more than 300 high-density cells effectively during hundreds of video frames, and the accuracy is up to 93.4%.

Key words: cell, tracking, group behavior, GPU acceleration, kinesiology, morphology

摘要： 提出一种基于关键帧的高密度细胞群体双向跟踪算法，手动标记中间帧的每只细胞，利用细胞纹理特征跟踪每只个体的运动。为解决重叠和遮挡问题，采用细胞运动学和形态学信息融合技术对其运动进行约束。针对细胞过分拥挤导致的轨迹片段，将视频的首尾分别作为关键帧，通过重新跟踪丢失细胞的方式把断裂的轨迹补充完整，并使用GPU对跟踪算法进行加速。实验结果显示，该算法能够在数百帧视频中，有效地跟踪300多只高密度不规则运动的细胞，其跟踪准确率高达93.4%。

关键词: 细胞, 跟踪, 群体行为, GPU加速, 运动学, 形态学

CLC Number: 

• TP18