基于多模态背景模型和霍夫森林的红外目标跟踪

Infrared target tracking based on multimodal background model and Hough forest

摘要:

针对红外目标跟踪过程中目标和背景耦合性强和目标形变带来的表观特征变化的问题,提出了基于多模态背景模型和霍夫森林的红外目标跟踪算法,将目标和背景的特征信息整体融入跟踪框架,并引入霍夫森林对目标部件表征信息进行训练更新；通过特征描述和模式列表建立多模态背景模型,在观测序列中记录匹配的背景模式列表,根据匹配结果进行权重和模态参数更新；以背景建模获得的区域信息为采样基础,将前景区域和背景区域分别作为正负样本集代入决策树进行训练；根据叶节点存储的图像块信息对目标位置进行投票决策,从图像特征空间映射到霍夫参数空间,生成目标区域概率分布图,获取高置信度区作为目标跟踪区。在红外公开测试集上实验结果表明,在背景杂波和目标形变对跟踪造成干扰的情况下,所提算法仍能保持跟踪的稳定性。

Abstract:

Aiming at the problems of strong coupling between the target and the background and the change of the apparent features caused by the target deformation in infrared target tracking, an algorithm based on multi-modal background model and Hough forest was proposed. Both the target and background feature information were integrated into the tracking framework, and Hough forest was utilized for training and updating the target component characterization information. The multi-modal background model was built upon the feature descriptions and the pattern lists. Along with the matching background pattern lists recording in the observed image sequence, the modal parameters were updated according to the matching results. After obtaining the regional information from the background modeling, the foreground and the background were respectively taken as the positive and negative sample sets for the training in the decision trees. The decision of target’s position was carried out through Hough voting operation according to the image blocks’ information stored by the leaf nodes. The images’ feature space was mapped to the Hough parameter space to generate the probability distribution map of the target, and the high confidence area was determined as the tracking position. Experimental results on the public infrared datasets show that the proposed algorithm can keep track the target stably under the circumstance that background clutter and target deformation interfere with the tracking process.