基于单目标模型扩展的多目标跟踪算法  

作　　者：陈嘉炜 

机构地区：[1]北方工业大学电气与控制工程学院,北京100144

出　　处：《工业控制计算机》2022年第7期82-84,共3页Industrial Control Computer

摘　　要：为了解决传统单目标跟踪方法在直接扩展到多目标场景中的精度和速度问题,提出了一种融入亲和性模型的多任务学习模型,融合了轻量化网络部分结构,实现了参数量和计算量的优化。针对该类方法自身局限性导致的轨迹错位和漂移等跟踪效果,在其中加入度量学习模块增强类内目标的区分性,引入了新型的融合注意力机制以提升跟踪的稳定性和增强特征学习,而得到最优跟踪结果。该模型在多目标挑战数据集MOT16上进行了测试,算法在维持不错的精度和鲁棒性的情况下,实现了跟踪速度上的稳定提升。In order to solve the accuracy and speed problems of the traditional single-target tracking method in the direct extension of multi-objective scenarios,a multi-task learning model integrated into the affinity model is proposed,which integrates the lightweight network part structure to achieve the optimization of parameter quantity and calculation amount,and for the tracking effects such as trajectory misalignment and drift caused by the limitations of the method itself,the measurement learning module is added to enhance the differentiation of targets in the class,and a new type of fusion attention mechanism is introduced to improve the stability of tracking and enhance feature learning,to get optimal tracking results.The thesis model is tested on the multi-objective challenge dataset MOT16 and the algorithm achieved a stable improvement in tracking speed while maintaining good accuracy and robustness.

关 键 词：多任务学习 三重损失函数 孪生网络 融合注意力 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]