多摄像机下的目标图像跟踪关联算法研究

Under the Camera Target Track Correlation Algorithm Research

传统算法在进行目标图像跟踪时,对系统模型和后验分布的限制较为严格,只能处理线性、高斯、单模态的情况,而图像跟踪应用中的后验概率呈非线性、非高斯、多模态的形式,且由于传统算法的计算量巨大、采样效果非常低,因此无法满足实际跟踪场景复杂性的需要。提出一种多摄像机下的目标图像跟踪关联算法。提取在目标跟踪过程中图像的主要特征,组建图像特征集,通过计算目标图像特征的均布差值来划分目标与背景间的差别,将目标范围约束在对应核的空间中,并在进行目标图像识别时融合EM算法,依据组建目标图像投影熵特征的分布模型,获取目标图像中每个目标特征与其相对应的混合高斯函数的Mahalanobis间距,得到目标图像特征的所属类别,进而完成精确的目标图像跟踪。实验结果证明,多摄像机下的目标跟踪关联算法精确度高,效率高。

Traditional algorithm in the target image tracking, because of the kalman filtering for system model and the posterior distribution of more stringent restrictions, can only deal with linear, gaussian, single mode, and the posterior probability image tracking application is nonlinear, non-gaussian, the form of the modal, so we can not meet the requirements of target tracking; Particle filtering for system model has no special requirements, and can maintain state of muhimodal distribution, less susceptible to noise, but the large amount of calculation, the effect of sampling is very low, cannot satisfy the needs of tracking scene complexity. For this, put forward a target image under multiple cameras to track correlation algorithm. To extract the main characteristics in the process of target tracking, the forma- tion of image feature set, through the calculation of uniform difference of target image characteristics to differentiate the differences between target and background, the target range constraint in the corresponding space of nuclear, and EM algorithm in the target image recognition fusion, according to set up the distribution of target image projection en- tropy model, to get each target in the target image characteristics and its corresponding mixed gaussian function of Mahalanobis distance, the class of image features, the target is obtained and then achieve the aim of accurate image tracking. Simulation experiment proves that the target track correlation algorithm under multiple cameras high preci- sion, high efficiency.