基于复杂场景下特征融合与估计的目标跟踪算法

Object Tracking Algorithm Based on Feature Fusion and Evaluation in Complicated Environment

摘要:

针对图像跟踪过程中的背景干扰、光照条件变化与目标特性漂移等问题,提出一种粒子滤波框架下灰度和纹理特征融合与估计的目标跟踪算法.首先建立基于灰度直方图和边缘梯度直方图的特征集合,以增加目标描述的信息量和稳定性;然后在跟踪过程中依据每个特征对目标与背景的区分度赋予其相应的权值,突出贡献较大的特征,并将该加权系数用于目标和候选目标的相似性度量;最后,为了克服复杂场景对单幅图像中特征可信度的影响,将跟踪的连续性引入权值的自适应调整过程,利用当前帧的特征区分度和前一帧的特征权值对当前帧的特征权值进行估计,从而实现特征权值的在线更新.实验结果表明,通过对特征的融合及其权值的在线估计,实现了复杂场景下的稳定目标跟踪.

Abstract:

Aiming at background interference,the variance of light and the drift of targets in the visual tracking process,an object tracking algorithm combining gray and contour features under the framework of particle filter was proposed.Firstly,a combined feature set with histogram of gray (HG) and histogram of gradient orientation (HOG) was established to enrich the target information and improve the stability.Secondly,in the process of tracking,each feature is given a confidence according to the ability of discriminating the target and background,which is regarded as a weight and applied to measure the similarity between the target and the candidate.Lastly,the temporal consistency was introduced into the online process of adjusting the confidence adaptively,in which the discriminative ability of the feature in current frame and the corresponding confidence in past frame are modeled to evaluate the confidence of the target in current frame,so as to avoid the effectiveness of complex environment on the confidence in a single image.The experimental results indicate that the proposed algorithm has a well-behaved performance in complex scene with combined features and the evaluation of the feature confidence online.