基于头肩模型的人体识别方法

Human Recognition Approach Based on Head-shoulder Model

针对移动机器人跟踪特定人体的要求,提出了一种基于头肩模型的人体识别方法;首先从人体检后得到的图像中提取所有的人体头肩模型;接着提取各头肩模型的降维加权的Hu不变矩作为特征向量;然后根据一定的阈值将各头肩模型分类为正背面或侧面;最后采用正背面或侧面K最近邻(K NearestNeighbor,KNN)分类器判断哪个头肩模型属于移动机器人需要跟踪的人体;实验结果表明本方法具有较高的平均识别准确率98.3%,且满足实时性的要求。

According to the requirements of tracking desired human for mobile robots, a human recognition approach which is based on the head-- shoulder model is proposed. Firstly, the human head-- shoulder models are extracted from the image obtained by human detection. Next, dimensionality reduction and weighted Hu moment invariants of the head--shoulder models are extracted as the feature vectors. Then, the head--shoulder models are identified as the front--back or profile models according to certain thresholds. Finally, the front--back or pro- file K Nearest Neighbor （KNN） classifier is used to determine which head--shoulder model belongs to the desired human, who needs to be tracked by a mobile robot. The experimental results show that the proposed approach has high average recognition accuracy 98. 3% and is provided with real--time performance.