基于加权在线样本更新的目标长时跟踪方法

针对在长时跟踪过程中因丢失视野导致目标跟踪失败的问题,提出了基于加权在线样本更新的目标长时跟踪方法。 首先,使用 ResNet50 网络提取目标深度特征并增强初始帧样本优化目标模型,提高初始帧样本权重影响;然后,利用目标模型对测试帧样本进行分类,并采用置信度分值加权在线学习样本以增强样本质量,提升模型的分类效果;其次,使用置信度分值判别目标状态并跟踪定位目标,目标丢失时使用时空约束搜索在丢失处自适应扩展区域并随机搜索目标,同时利用在线学习快速优化目标模型,增强其对目标的搜索能力;最后,针对搜索过程设计一种自适应阈值判别方法,充分利用图像背景信息,将目标丢失时背景置信度分值作为判别阈值,降低搜索过程中相似背景的影响以准确找回目标。 使用 LTB50 数据集进行实验验证,成功率和跟踪 F-score 分别为 66. 1% 和 64. 4% ,优于其他方法;在四足移动机器人平台上进行真实场景实验,目标完全遮挡和视野外两种情况下成功率分别为 87. 8% 和 85. 8% ,证明了方法的有效性。

Target long-term tracking method based on weighted online sample update

A long-term tracking method based on the weighted online sample updating is proposed to address the problem of trackingfailure caused by target loss during long-term tracking. First, the ResNet50 network is used to extract the deep features of the target andenhance the initial frame sample to optimize the target model, which could improve the influence of the initial frame sample weight.Then, the target model is used to classify the test frame sample, and the confidence score is used to weight the online learning samplesto enhance their quality and improve the classification performance of the model. Secondly, the target state is determined by theconfidence score, and the target is tracked and located. When the target is lost, a spatiotemporal constraint search is used to adaptivelyexpand the search area at the loss point and randomly search for the target, while utilizing online learning to quickly optimize the targetmodel and enhance its search ability. Finally, an adaptive threshold discrimination method is proposed for the search process, fullyutilizing the image background information, using the background confidence score when the target is lost as the discrimination threshold,reducing the influence of similar backgrounds in the search process to accurately retrieve the target. Experiments on the LTB50 datasetshow a success rate of 66. 1% and a tracking F-score of 64. 4% , outperforming other methods. Real-world experiments on a quadrupedrobot platform achieved success rates of 87. 8% and 85. 8% under full occlusion and out-of-view scenarios, respectively. Theeffectiveness of the proposed method is evaluated.