基于DF-Track的水下鱼体跟踪方法

Method of underwater fish body tracking based on DF-Track

鱼类跟踪是分析鱼类行为、评估其健康水平的关键步骤。然而,由于真实水下养殖鱼群具有运动非线性、高外观相似度、鱼体互相遮挡、特征信息损失严重等特点,多目标跟踪鱼类是一项非常具有挑战性的任务。针对水下鱼体反射产生的伪影以及鱼体运动非线性、相互遮挡导致跟踪轨迹碎片化的问题,提出一种水下多鱼跟踪模型DF-Track。该模型采用基于检测的跟踪(TBD)范式,首先在YOLOv8的C2f结构中引入聚合感知注意力机制(APFA),提高在前向过程中提取图像特征的能力;然后使用SDI多层次特征融合模块对YOLOv8中的feature fusion部分进行重设计,减少特征融合阶段不同层级特征信息的冲突问题;最后提出一种优化轨迹管理的跟踪模型DF-Track,并引入Focal-EIoU补偿匹配空间中的运动估计偏差,平衡几何一致性。实验结果表明:与原YOLOv8相比,所提算法的精确率提高了1.7%,平均精度均值提高了2.1%;DF-Track与其他MOT跟踪算法相比,HOTA达到70.9%,MOTA达到91.9%,IDF1达到80.4%。证明DF-Track模型在水下鱼类跟踪任务中具有较好的性能。

Fish tracking is a crucial step in analyzing fish behavior and assessing their health status.However,due to the characteristics of real underwater aquaculture fish schools,such as nonlinear motion,high appearance similarity,mutual occlusion between fish bodies,and severe loss of feature information,multi-target tracking of fish is a highly challenging task.To address the issues of artifacts caused by underwater fish body reflection,as well as the fragmentation of tracking trajectories due to nonlinear fish motion and mutual occlusion,an underwater multi-fish tracking model DF-Track is proposed.In this model,the tracking by detection(TBD)paradigm based on detection is adopted,and the aggregated pixel-focused attention(APFA)mechanism is introduced into the C2f structure of YOLOv8 to improve the ability to extract image features during the forward process.The SDI multi-level feature fusion module is used to redesign the feature fusion section in YOLOv8,reducing the conflicts in feature information at different levels during the feature fusion stage.Then,an optimized trajectory management tracking model,DF-Track,is proposed,and Focal-EIoU is introduced to compensate for motion estimation deviations in the matching space,so as to balance geometric consistency.The experimental results show that,in comparison with original YOLOv8,the accuracy of the proposed algorithm is increased by 1.7%,and the average accuracy is increased by 2.1%;in comparison with other MOT tracking algorithms,DF-Track can realize 70.9% in HOTA,91.9% in MOTA and 80.4% in IDF1.It proves that DF-Track model has better performance in underwater fish tracking task.