用于实时跟踪的模板匹配神经网络算法

研究复杂背景下运动目标的识别和跟踪.提出了一种新的基于模板匹配的神经网络结构,将模板与跟踪窗内待匹配区域的像素按环形排列,分别作为神经网络的阈值和输入,选择跟踪窗内与模板相对应的各环差值均较小的区域作为识别结果.由于模板匹配过程中像素按环形排列,因此对于目标的平移和旋转均具有不变性,同时,算法计算量比最小绝对差累加和算法略小.将该算法应用到实时跟踪系统中,实验结果表明该算法可满足跟踪系统实时性要求,验证了算法的有效性.

Neural Network Based on Template Matching and Its Application in Real-Time Target Tracking

A novel neural network based on template matching is proposed for target recognition and tracking. Pixels in template and tracking window are arranged to form into loops, which are set to the threshold and the input of the neural network. The region where the sum of pixels in every loop is near the one in the template is chosen as the recognition result. Because the pixels are arranged into loops, the algorithm is of translation and rotation invariance. Furthermore, the calculated amount of the algorithm is somewhat less than that of MAD (minimum absolute difference). The algorithm can be applied to the real-time tracing system. The experimental results prove that the algorithm can satisfy the real-time need of the tracing system.