为了满足某些算力受到限制的应用场景的长时跟踪需求, 如以C64x+ DSP为核心的嵌入式系统, 提出了一种由连续跟踪环节与目标检测环节两部分构成的低时间复杂度长时跟踪算法, 连续跟踪环节基于自适应更新的时空上下文算法(STC), 目标检测环节使用归一化互相关匹配算法。在没有目标出视场、目标快速移动等特殊跟踪场景时, 连续跟踪环节输出跟踪结果, 在跟踪失败后, 目标检测环节对全幅图像进行处理, 只要目标出现在图像中, 便可以重新锁定目标。经实验验证, 目标检测环节可以在目标出现后准确检测到目标, 满足了长时跟踪的要求。同时, 目标检测环节在跟踪不可靠时的辅助定位也提升了连续跟踪环节的鲁棒性, 使用OTB2013数据集测试, 本算法的精确度较STC算法提升了4.95%。

In order to meet the long-term tracking requirements of platforms with weak computing power, such as the embedded system with C64x+ DSP as the computing core, a long-term tracking algorithm with low time complexity was proposed, which consisted of two parts. One part was the continuous tracking part and the other part was the target detection part. The continuous tracking part was based on adaptive update spatio-temporal context algorithm(STC), and the target detection part used normalized cross-correlation matching algorithm. If there were no special tracking scenarios such as field of view and fast moving target, the continuous tracking part outputted tracking results. After tracking failure, the whole image was processed in the target detection part, and the target was re-locked as long as the target appeared in the image again. Experiments show that target detection can accurately detect the target after it appeared, which meets the requirements of long-term tracking. At the same time, the robust of continuous tracking is also improved because the target location is redefined by the target detection part, when the result is not reliable. Using OTB2013 data set to test, the accuracy of this algorithm is 4.95% higher than that of STC algorithm.