随着5G、大数据、人工智能等高新科技不断发展，目标跟踪技术已经成为计算机视觉的一个重要研究热点，为后续的目标智能行为分析工作提供有效帮助，广泛应用在安防监控、军事制导、智能交通以及视觉导航等领域。

相关滤波目标跟踪算法具有较强的稳定性和鲁棒性，是近年来判别式跟踪算法的重要里程碑。在实际应用环境中，跟踪算法面临着场景多样性和目标尺度变化复杂等难点，多尺度核相关目标跟踪算法能有效地提高跟踪性能。目标尺度估计的准确度对跟踪性能具有重要影响，但是随着检测尺度数量的增加，其跟踪时间效率会严重下降，从而导致了多尺度跟踪算法的非实时性。
    本论文采用两种改进策略充分利用跟踪结果的反馈信息进行分类处理，提出了基于响应判别的多尺度相关滤波跟踪算法。改进策略1基于跟踪目标多分辨率分段预处理方法，能有效提高低分辨率属性视频的跟踪性能。改进策略2基于响应判别的多尺度检测方法，通过最大响应峰值的判别分类处理，对于目标跟踪较为准确的视频帧不进行多尺度检测，从而节省大量运算时间；对于目标跟踪性能不准确的视频帧进行多尺度检测，从而修正结果提高跟踪性能。
    在OTB50、OTB100、TC128和UAV123数据集上对比当前先进的多尺度相关滤波目标跟踪算法，实验结果表明本文改进的算法不但能够有效地提高跟踪速度，实现实时目标跟踪，而且跟踪性能也有显著提升。在OTB100视频库中，本文改进算法的DP精确度平均得分比SAMF、SAMF_CA、fDSST、DSST、KCF算法分别提高了4.4%、0.3%、7.1%、10.1%和10%；AUC成功率平均得分分别提高了2.7%、0.5%、2.5%、5.7%和10.3%；本文改进算法的目标跟踪速度分别是SAMF、SAMF_CA算法的1.65倍和1.93倍。

With the continuous development of high-tech such as 5G, big data and artificial intelligence, target tracking technology has become an important research hotspot in computer vision, providing effective help for subsequent target intelligent behavior analysis. It is widely used in security monitoring, military guidance, intelligence transportation and visual navigation.

Correlation filters target tracking algorithm has strong stability and robustness, which is an important milestone of discriminant tracking algorithm in recent years. In the actual application environment, tracking algorithms are faced with the difficulties of scene diversity and complex changes of target scale. Multi-scale correlation filters target tracking algorithms can effectively improve tracking performance. The accuracy of target scale estimation has an important impact on tracking performance, but as the number of detection scales increases, its tracking time efficiency will seriously decrease, which has bad effect on real-time performance of multi-scale tracking algorithms.

This paper uses two improved strategies to make full use of the feedback information of the tracking results for classification processing, and proposes a multi-scale correlation filtering tracking algorithm based on response discrimination. Improvement strategy one: Multi-resolution segmentation preprocessing method based on tracking target can effectively improve the tracking performance of low-resolution attribute video. Improvement strategy two: Multi-scale detection method based on response discrimination, through discriminant classification processing of the maximum response peak, multi-scale detection is not performed for video frames with more accurate target tracking, there by saving a lot of computing time. For video frames with in accurate target tracking performance, multi-scale detection is performed to correct the results and improve tracking performance.

Comparing the current advanced multi-scale correlation filters target tracking algorithms on the OTB50, OTB100, TC128 and UAV123 data sets, the experimental results show that the improved algorithm in this paper can not only effectively increase the tracking speed and achieve real-time target tracking, but also significantly improve the tracking performance. In the OTB100 video library, the DP tracking accuracy average score of the improved algorithm in this paper is 4.4%, 0.3%, 7.1%, 10.1%, and 10% higher than SAMF, SAMF_CA, fDSST, DSST, and KCF respectively; the average score of the AUC success rate is 2.7%, 0.5%, 2.5%, 5.7% and 10.3% higher; the tracking speed of the improved algorithm in this paper is 1.65 times and 1.93 times of SAMF and SAMF_CA algorithms respectively.