基于改进FairMOT特征解耦的多目标跟踪算法

Multi-Object Tracking Algorithm Based on Improved FairMOT Feature Decoupling

联合检测和重识别跟踪模型(Joint-Detection-and-Embedding Models, JDE)的两个子任务所需要的特征存在矛盾,通过目标中心点提取重识别特征的方式难以得到遮挡目标的有效特征,这导致在复杂环境下模型提取的目标重识别特征可靠性下降,造成数据关联错误。针对目标检测和重识别任务间的矛盾问题,文中基于FairMOT跟踪算法提出了一种特征解耦模块。该模块使用协调注意力(Coordinate Attention, CA)将骨干网输出的多尺度特征图进行初步解耦,然后以自底向上的方式融合不同分辨率的重识别特征图。为了提取遮挡目标的有效信息,文中提出一种根据目标可视度调整高斯核方差的策略,用于构建目标中心点监督热图,加大训练时对遮挡目标及其周围区域的关注。最后在MOT17数据集上对所提算法进行了测试,实验结果验证了各模块的有效性,表明了算法能够有效应对遮挡,实现稳定跟踪。

The features required for the two sub-tasks of joint object detection and re-identification mul-ti-object tracking algorithm (Joint-Detection-and-Embedding Models, JDE) are contradictory. It is difficult to extract the effective features of occluded objects by extracting object re-identification features through the object center point. This leads to unreliable object re-identification features extracted by the model in complex environments, resulting in data association errors. A feature decoupling module is proposed based on the FairMOT tracking algorithm, aiming at the contradiction between object detection and re-identification tasks. This module uses coordinate attention to initially decouple the multi-scale feature maps output by the backbone network, and then fuses re-identification feature maps of different resolutions in a bottom-up manner. In order to extract the effective information of the occluded object, a strategy of adjusting the variance of the Gaussian kernel according to the visibility of the object is proposed, which is used to construct a supervised heat map of the object center point, and pay more attention to the occluded object and its surrounding areas during training. The proposed algorithm is tested on the MOT17 dataset, and the experimental results verify the effectiveness of each module, indicating that the algorithm can effectively deal with occlusion and achieve stable tracking.