提出了一种基于自适应特征融合和自适应模型更新的相关滤波跟踪算法(CFT)。该算法在跟踪的训练阶段利用损失函数计算特征的自适应权重, 在检测阶段对不同特征的响应图进行加权求和, 从而实现了响应图层面的自适应特征融合。设计了自适应的模型更新策略, 采用响应图的峰值旁瓣比判断是否发生遮挡或错误跟踪, 据此决定是否在当前帧更新目标模型。在11个视频序列上对所提算法进行了实验, 验证了所采用的自适应特征融合策略和自适应模型更新策略的有效性。与多个传统的采用单特征的相关滤波跟踪算法进行了比较, 结果显示, 所提算法的跟踪精度和成功率典型值分别提升了18.2%和11.5%。实验结果验证了特征融合和自适应模型更新对跟踪算法的改进具有指导意义。

A Correlation Filter based Tracking(CFT) method based on adaptive feature fusing and model adaptive updating is proposed. The proposed method computes the adaptive feature weights by using the loss function of a filter model in training stage, and performs the weighted summing for response maps with multiple features in detection stage. Then it realizes the adaptive feature fusing of response maps. Furthermore, an adaptive model updating strategy is designed to estimate the occlusions and wrong tracking results by utilizing the Peak to Sidelobe Ratio (PSR) of the response maps. The model is updated only when the PSR is small enough. The proposed method is tested on 11 video sequences to verify the validity of the proposed adaptive feature fusing and model adaptive updating strategies. Comparing with other CFT methods by using single feature, the proposed tracker has significantly improvements, by 18.2% in representative precision score and 11.5% in representative success score. Experimental results demonstrate that the adaptive feature fusing strategy and model updating strategy are all effective for improving tracking performance.