由于运动模型采用线性假设,使得多目标跟踪(multiple object tracking,MOT)在动态场景下容易受到相机运动和随机抖动的影响,导致跟踪错误.为解决上述问题,设计了一种相机运动感知滤波多目标跟踪器(camera motion aware filter multi-object tracker,CMAFT).首先提出一种新模型,将相机运动估计和单目标跟踪(single object tracking,SOT)的区域搜索特性相结合,以补偿由相机运动引起的偏移并提高预测的准确度;其次针对该模型提出一个改进的级联匹配方法,通过融合SOT预测以处理不同目标间的相互遮挡和身份切换问题;最后在MOT17数据集上进行实验以验证提出方法的有效性.
Multiple object tracking(MOT)is susceptible to the motion and random jitter of the camera in dynamic scenes due to the linear assumption of the motion model,leading to tracking errors.A camera motion aware filter multi-object tracker(CMAFT)is designed to solve the above problems.Firstly,a new model combines camera motion estimation with the regional search characteristics of single object tracking(SOT)to compensate for the drift caused by camera motion and improve the accuracy of prediction.Secondly,an improved cascade matching method for this model is proposed to deal with the mutual occlusion and identity switching between different objects by fusing the SOT prediction.Finally,experiments are run on the MOT17 dataset to verify the effectiveness of the proposed method.
陈健超;奚峥皓;刘翔
上海工程技术大学 电子电气工程学院,上海 201620
计算机与自动化
多目标跟踪(MOT);相机运动估计;单目标跟踪(SOT)
multiple object tracking(MOT);camera motion estimation;single object tracking(SOT)
《计算机工程与应用》 2024 (010)
285-291 / 7
国家自然科学基金(12104289).
10.3778/j.issn.1002-8331.2303-0006
评论