基于深度学习的视频序列运动目标自适应跟踪

Adaptive tracking of moving targets in video sequences based on deep learning

摘要:

针对视频序列中外观变化、背景杂波和严重遮挡等因素导致的目标跟踪精确度低的问题，提出一种新型的双阶段自适应跟踪模型。该模型包含目标检测和边界框估计2个阶段：在目标检测阶段，模型对目标进行粗略定位；在边界框估计阶段，精确确定目标位置。为应对视频场景复杂性及小目标跟踪的挑战，采用了多特征融合技术构建丰富的目标表示。实验结果表明，与在线和实时跟踪(SORT)、Tracktor++、FairMOT、Transformer等模型相比，本模型表现出最优的综合性能，有效平衡了计算速度与跟踪精确度之间的关系，展现出良好的应用潜力。

Abstract:

In response to the issues of low tracking accuracy in video sequences due to factors such as appearance changes, background clutter, and severe occlusions, a novel two-stage adaptive tracking model is proposed. This model includes two phases: target detection and bounding box estimation. In the target detection phase,the model roughly locates the target; in the bounding box estimation phase, the exact position of the target is determined. To address the complexity of video scenes and the challenges of tracking small targets, multi-feature fusion technology is employed to construct a rich target representation. Experimental results show that compared with models such as Simple Online and Realtime Tracking(SORT), Tracktor++, FairMOT, and Transformer, this model demonstrates the best overall performance, effectively balancing the relationship between computational speed and tracking accuracy, and showing good potential for application.