复杂场景下基于自适应多特征融合的跟踪算法

在复杂的场景下，单特征对目标描述不够充分，很难稳健地跟踪目标，针对这个问题，提出了一个基于自适应多特征融合的粒子滤波跟踪算法。该算法采用灰度和边缘特征表示目标，从目标观测似然模型构建的角度融合两种特征，利用粒子似然分布的香农熵动态地评价特征的可靠性，进而确定特征融合权重，以提高算法对场景的适应能力；同时,改进了线性加权的模型更新策略，通过对加权系数的在线调整来抑制模型漂移。实验表明，本文算法可以实现部分遮挡和背景干扰等复杂场景下的跟踪。

Tracking Algorithm Based on Adaptive Multi-features Fusion Under Complex Scenarios

It's difficult to track object stably because of the shortcomings of single feature under complex scenarios. Aming at this problem, an particle filter tracking algorithm based on adaptive multi-features fusion is proposed. Target is represented by the gray and edge, this two features are fused from the perspective of target observation likelihood model construction, The proposed algorithm dynamically assesses feature's reliability by the Shannon entropy of particles' likelihood distribution, then determines the feature's fusion weight with respect to it's discriminability. Simultaneously, we improve the linear weighted model update strategy by adjusting the weighting coefficient on line,which suppresses model drift. Experiments show this algorithm can achieve tracking under complex scenarios such as partial occlusion and background interference.