中文摘要：

为解决目标外形、姿态变化以及被遮挡的难点，对压缩跟踪算法进行改进：以Kalman预测位置为中心，采用由粗到精的搜索策略，快速准确地找到具有最大分类分数的目标位置；根据目标在每一帧最大分类分数的变化规律进行粗判定，再利用目标和模板的相似度进行精判定以判断遮挡的开始；利用Kalman预测器预测目标被遮挡时的位置，利用基于MCD距离的模板匹配检出脱离遮挡后的目标。使用浮点转定点等策略将算法在DSP上优化实现。实验结果表明：跟踪效果稳定，在目标外形、姿态缓慢变化以及被遮挡时能准确地跟踪目标。对640×480像素的视频，跟踪框为64×64像素时，被遮挡目标跟踪速度达到38．5毫秒/帧，满足实时性要求。

英文摘要：

To solve the problems of shape change, pose variation and occlusion in target tracking, Compressive Tracking algorithm is improved by ways below：taking the Kalman predictor’s output as the center of search area, a coarse-to-fine search strategy is used to quickly and accurately find the target position with the maximum classification score. We first roughly determine whether occlusion happens according to the regular pattern of variation of the maximum classification score of each frame. Then we determine whether occlusion happens exactly by computing the similarity between object and template. The Kalman predictor offers the predicted location of the object when occlusion happens. The judgment of the end of occlusion and re-detection of target is completed by template matching based on the MCD method. Real-time target tracking is realized based on the improved CT algorithm in DSP with optimization tactics such as changing float-point computation to fixed point computation. Experiments show that the result of target tracking is stable. The target can be tracked accurately when slow shape change, slow pose variation and occlusion happen. When the input video’s size is 640 ×480 pixels and the tracking rectangle＇ size is 64×64 pixel, the occluded target is tracked with the speed up to 38.5ms per frame, which meets the real-time requirements.