一种结合多特征的实时物体识别系统

Real-time Object Recognition System Combining Multiple Features

物体识别是众多智能系统应用的一项基本功能.针对物体识别的实时性和准确性要求,提出结合多种特征的物体识别方法,在此基础上实现基于多特征融合并支持动态阈值调整的物体识别系统.该系统首先采用快速模板匹配的方法在RGB-D图像上检测目标物体,再利用颜色特征过滤误识别,最后采用局部特征匹配的方法在高分辨率图像上精确识别物体.通过动态地自适应调整这三步的阈值参数,能够在保证实时性的情况下显著提高识别率.实验结果表明,在复杂多变的应用环境下,该系统能以低于0.2秒的平均识别时间得到高识别率,满足服务机器人对物体检测与识别的实时性和准确性的要求,并多次在实际应用中展现出实用性和可靠性.

Object recognition is a basic functionality for many intelligent systems. For satisfying its real-time and accuracy require- ments, we propose an object recognition method that supports multi-feature fusion and dynamic thresholds adjustment. Our method first employs a fast template matching technique to detect targets in RGB-D images,and then applies the color feature to filter false detec- tion results. Lastly, local feature matching approach is utilized to recognize object more accurately in images with higher resolution. By adjusting threshold parameters in these three steps dynamically and self-adaptively, recognition accuracy can be enhanced while real- time performance can be achieved. Experiments show that our method can achieve high recognition accuracy with average recognition time less than 0. 2s. Our method meets the object recognition requirements for robots in complex environments and has been demon- strated to be practical and reliable for many times in practice.