融合运动状态信息的高速相关滤波跟踪算法

为解决相关滤波(Discriminative Correlation Filter,DCF)算法在快速运动、遮挡、尺度变化等复杂情景下的跟踪失败问题,提出一种融合运动状态信息的高速相关滤波目标跟踪算法.在传统DCF算法基础上做出以下改进:(1)在跟踪框架中融入卡尔曼(Kalman)滤波器,利用目标运动状态信息对预测运动轨迹进行修正,以解决目标复杂运动时易跟丢问题,提高跟踪精度;(2)训练一个独立的尺度相关滤波器进行目标尺度预测,并利用主成分分析法(Principal Component Analysis,PCA)进行特征降维处理,提高跟踪速度;(3)提出一种高置信度更新策略判断是否对位置滤波器进行模板更新,以及是否采用Kalman滤波器预测位置作为目标位置.最后在OTB-100数据集上进行算法测试,提出算法平均精度与成功率分别达到74.8%与69.8%,平均帧率为84.37帧/s.相较其他几种主流算法,本文算法有效提高跟踪性能,并保证了跟踪速度,满足实时性要求,在遮挡、背景模糊、运动模糊等复杂情况下能够保持良好的跟踪效果.

High Speed Correlation Filter Tracking Algorithm Integrating Motion State Information

In order to solve the problem of tracking failure caused by complex scenarios such as fast motion, occlusion and scale variation, a high-speed correlation filtering target tracking algorithm integrating motion state information is proposed. This paper makes three improvements based on the traditional Discriminative Correlation Filter: (1) The Kalman filter is added to the tracking process to modify the predicted position by using the motion state information, so as to deal with the tracking failure caused by fast motion and improve the tracking accuracy; (2) A separate filter for scale estimation is learned and the PCA method for dimension reduction of features is used to improve the tracking speed. (3) A high-confidence update strategy is proposed to determine whether the position filter is updated and whether the predicted position is transferred to Kalman filter for correction. The algorithm is tested on OTB-100 platform with several state-of-the-art tracking algorithms. Experiments show that our algorithm''s average precision and success rate can reach 74.8% and 69.8%, respectively, and the average speed is 84.37 frames per second. Compared with other algorithms, the proposed algorithm can effectively improve the tracking performance, guarantee the tracking speed, and keep good tracking effect under complex conditions such as occlusion, ambiguous background and fast motion.