针对由于目标频繁遮挡、不规则运动导致的外观特征不可靠和运动特征难以获取的问题,提出一种基于膨胀交并比区域(dilatation intersection over union,DIOU)匹配和自适应轨迹管理策略的多目标跟踪算法.DIOU模块通过膨胀匹配区域,提升轨迹级联匹配的精度.自适应轨迹管理策略利用目标检测置信度动态调整轨迹生命周期,显著减少了异常跟踪和身份跳变.在公开数据集MOT17、MOT20和DanceTrack上进行了验证与测试,其在测试集上的高阶跟踪精度平均提升了 2.4％,实验结果证明了所提方法的有效性.
A Multi-object Tracking Method Based on Dilatation Region Matching and Adaptive Trajectory Management Strategy
Objectives:Multi-object tracking(MOT)is a pivotal research area within the computer vision domain.Despite significant strides in MOT research,the field continues to grapple with formidable chal-lenges:Indistinct appearance attributes of objects,objects exhibit irregular motion,anomalies in tracking arising from rigid trajectory lifecycle management strategies.These elements substantially undermine the precision and robustness of multi-object tracking endeavors.Methods:In response to these challenges,we present an advanced multi-object tracking algorithm that integrates dilatation intersection over union(DIOU)matching with an adaptive trajectory management approach.Initially,we introduce a metric based on a refined DIOU area for the primary matching between active trajectories and high-confidence detec-tions,thereby improving the direct matching performance for high-quality detection boxes.Subsequently,for the re-matching of active trajectories with low-confidence detections,we implement a metric centered on a moderately dilated DIOU area,enhancing the tracking continuity of these detections.Furthermore,for reconnecting inactive trajectories with unmatched high-confidence detections,we employ a metric utilizing an extensively dilated DIOU area to bolster the probability of reactivating dormant trajectories.Lastly,an adaptive trajectory management strategy predicated on detection confidence scores is deployed to dynamical-ly modulate the lifespan of trajectories,thereby mitigating the incidence of tracking anomalies and identity switches induced by occlusions and misidentifications.Results:(1)The application of the DIOU-based matching framework has yielded 5.4％increase in HOTA(higher order tracking accuracy)and a 1.5％in-crease in MOTA(multiple object tracking accuracy)on the DanceTrack dataset,corroborating the meth-od's efficacy in densely populated scenes and complex motion environments.(2)The implementation of the adaptive trajectory management module has further resulted in 4.6％rise in HOTA,0.8％elevation in MOTA,and 2.1％improvement in IDF1(identification F-score)on the DanceTrack dataset,demon-strating its capacity to efficiently counteract the limitations of fixed lifecycle sensitivities to false detections and missed detections.Conclusions:Although the refinement of data association and trajectory management strategies has led to a surge in tracking accuracy,the layering of multiple strategies has introduced a trade-off with computational efficiency,curtailing the peak performance of the tracking system.
万方数据
维普
