基于压缩感知尺度自适应的多示例交通目标跟踪算法

针对大多数跟踪算法对车辆等交通目标在行驶过程中的尺度变化、姿态变化的适应性差及在跟踪过程中使用固定尺度的跟踪框，导致所构造的目标模板包含大量背景信息，引起跟踪漂移甚至丢失的问题，提出一种基于压缩感知理论与超像素目标性度量的尺度自适应多示例交通目标跟踪算法，该算法首先利用压缩感知理论对多示例学习中的特征维数进行降维，减少算法计算的复杂度。其次，采用超像素目标性度量进行局部尺度自适应调整，解决多示例跟踪算法中的尺度适应问题。此外，引入基于目标判别机制的分类器更新，利用连续帧中目标的相似性判断跟踪目标是否存在遮挡或漂移问题。依据目标判别的结果，实现变学习率的分类器参数更新。试验结果表明：该方法具有较高的跟踪精度和良好的跟踪鲁棒性，在车辆目标发生遮挡、尺度变化、三维旋转等情况时均能较好地跟踪目标，通过对不同的交通视频序列进行测试，算法的平均中心位置误差远小于对比算法，仅为3.92像素，其对比算法CT跟踪、MIL跟踪及WMIL跟踪算法的平均位置误差分别为56.96像素、35.36像素及58.54像素，平均重叠率达80.1%，较CT跟踪、MIL跟踪及WMIL跟踪算法分别高44.9%、45.3%和45%，满足智能交通监控的实际需求。

Traffic Target Tracking Algorithm Based on Scale Adaptive Multiple Instance Learning with Compressive Sensing

This paper aims to solve the problem that most existing tracking algorithms have poor adaptability in the process of tracking a traffic target with scale and pose changes and use a fixed-size tracking box to track the target, which results in the target template containing a lot of background information and causing tracking drift or failure. In this paper, a traffic target tracking algorithm based on scale adaptive multiple instance learning with compressive sensing theory and a superpixel objectness measure is proposed. Firstly, compressive sensing theory is used to reduce the feature dimension in multiple instance learning, thus reducing the computational complexity of the algorithm. Secondly, local scale adaptive adjustment is carried out by using the superpixel objectness measure to solve the scale adaptation problem in the multiple instance tracking algorithm. In addition, a target identification mechanism with variable learning rate is introduced to update the classifier parameters, and the occlusion or drift of the tracked target is judged by the similarity of the target in the continuous frame. According to the result of target identification, the classifier parameters are updated with the variable learning rate. The experimental results show that the proposed tracking algorithm has good robustness and high tracking accuracy under complex environments with vehicle occlusion, scale change, and three-dimensional rotation. Compared with CT, MIL, and WMIL trackers for the test traffic video sequences, the average center position error in test videos is far smaller, at only 3.92 pixels; the error in CT, MIL, and WMIL is 56.96 pixels, 35.36 pixels, and 58.54 pixels, respectively. The average overlap rate of the proposed method is 80.1%, while that of CT, MIL, and WMIL is 44.9%, 45.3%, and 45%, respectively, which meets the real application of intelligent traffic monitoring.