Abstract: At present, intelligent video surveillance has made a high demand for real-time, accuracy and robustness of video target tracking algorithm, but the existing algorithms cannot fully meet the application requirements. In this paper, a foreground classification algorithm based on Visual Background extractor（ViBe） is proposed to improve the speed of TLD detection target. The tracker in TLD framework is realized by Kernel Correlation Filter（KCF）, which improves the accuracy and robustness of the algorithm. To verify the feasibility of the proposed algorithm, OTB-2013 benchmark for video surveillance using is tested and compared with the other four representative tracking algorithms. The experimental results show that the improved TLD algorithm is superior to the contrast algorithm in the robustness and accuracy, and the processing speed can reach 40 frame/s. Compared with the standard TLD algorithm, the tracking distance is improved by 1.52 times and the success rate is improved by 1.2 times. Compared with the KCF algorithm, the tracking speed is improved by 2.7 times and the success rate is 2.04 times.

Key words: video surveillance, target tracking, tracking-learning-detection, kernelized correlation filter, visual background extractor

摘要： 目前智能视频监控对视频目标跟踪算法的实时性、准确性和鲁棒性都提出了很高的要求，而已有算法无法完全满足应用需求。在TLD（Tracking Learning Detector）框架下，提出一种基于视觉背景提取（Visual Background extractor，ViBe）的前景分类算法，提高了TLD算法检测目标的速度；用核相关滤波器（Kernelized Correlation Filters，KCF）实现了TLD框架中的跟踪器，提高了算法的精度及鲁棒性。采用OTB-2013评估基准中针对视频监控的视频序列进行测试，并与其他4种具有代表性跟踪算法进行了对比。测试结果表明：该算法的鲁棒性和准确性均优于对比算法，处理速度可达到40帧/s；相比于标准TLD算法，跟踪距离精度提高了1.52倍，成功率提高了1.2倍；相比于KCF算法，虽然跟踪速度有所下降，但跟踪距离精度提高了2.7倍，成功率提高了2.04倍。

关键词: 视频监控, 目标跟踪, 跟踪学习检测, 核相关滤波器, 视觉背景提取