Correlation filter tracking based on superpixel and multifeature fusion

For the correlation filtering(CF) tracking algorithm is not robust enough and cannot adapt to scale changes, target occlusion(OCC) and other complex interferences. We introduce a CF tracking algorithm based on superpixel and multifeature fusion(CFSMF). First, superpixel segmentation and clustering are performed for the target and its surrounding environment in the initial frame. Then, a target appearance is reconstructed through block segmentation-based overlapping analysis to remove redundant information. On this basis, the histogram of gradient(HOG) and HSI color features of the target sub-block are extracted to interact with their respective position filters. Accordingly, the target position is determined by the weighted fusion of the response values. In the scale prediction stage, we independently train a scale filter with a multiscale pyramid constructed at the estimated target location. The object scale is estimated in terms of the filter response, thereby enabling the tracking algorithm to adapt to the object scale change. Lastly, we introduce an OCC criterion for determining whether to update the model or not. Compared with the classical tracking algorithm kernelized correlation filters(KCF), the proposed algorithm boosts the tracking success rate by 20% and tracking accuracy by 15.9%. Our algorithm in this paper could track the target stably even when the target is occluded and its scale changes.     

基于改进高分辨率神经网络的多目标行人跟踪

针对行人多目标跟踪过程中目标被遮挡时产生的检测、跟踪失败问题,提出了一种改进型高分辨率神经网络作为检测网络。首先,为了增强网络对于行人目标的初始特征提取能力,在高分辨率神经网络的基础上,对网络的主干部分引入二代瓶颈残差块结构,提升感受野和特征表达力;其次,设计了添加二层高效通道注意力模块的残差检测块架构,并通过该架构替换了原有网络在多尺度信息交换阶段中的残差检测块,以提高了整个网络系统的测试性能;最后,通过选择适当的参数对网络进行了全面地训练,并通过多个测试集对算法测试。测试结果显示,本文算法相较于FairMOT在2DMOT15,MOT17,MOT20数据集上的跟踪准确度分别提升0.1%,1.6%,0.8%。本文算法可以良好地应用在目标较多且遮挡面积较大的特殊情景,同时对于较长时间视频序列的追踪稳定性也大大提高。     

基于卷积注意力模块和无锚框检测网络的行人跟踪算法

针对多目标跟踪过程中遮挡严重时的目标身份切换、跟踪轨迹中断等问题,该文提出一种基于卷积注意力模块(CBAM)和无锚框(anchor-free)检测网络的行人跟踪算法。首先,在高分辨率特征提取网络HrnetV2的基础上,对stem阶段引入注意力机制,以提取更具表达力的特征,从而加强对重识别分支的训练;其次,为了提高算法的运算速度,使检测和重识别分支共享特征权重且并行运行,同时减少头网络的卷积通道数以降低参数运算量;最后,设定合适的参数对网络进行充分的训练,并使用多个测试集对算法进行测试。实验结果表明,该文算法相较于FairMOT在2DMOT15, MOT17, MOT20数据集上的精确度分别提升1.1%, 1.1%, 0.2%,速度分别提升0.82, 0.88, 0.41 fps;相较于其他几种主流算法拥有最少的目标身份切换次数。该文算法能够更好地适用于遮挡严重的场景,实时性也有所提高。