论文部分内容阅读
针对人体关节点序列的连续行为识别问题,提出了一种基于BSCPs-RF(B-spline control points-random forest)的人体关节点信息行为识别与预测方法.首先采用局部线性回归与单帧关节点归一化法预处理关节点序列,以此消除抖动噪声、位移和尺度的影响;然后以B样条曲线控制点作为速度无关的关节点序列特征,并采用同步语音提示词法标注实时行为序列以提高样本采集效率;最后采用基于随机森林的行为识别与预测方法,并以集成学习方法优化多分类器组合以提高识别精度.实验分析了不同参数值对识别效果的影响,并分别在测试数据库MSR-Action3D以及RGB-D设备采集的实时数据集中进行测试.结果显示,MSR-Action3D测试结果优于部分先前方法,而实时数据测试中该方法具有很高的识别精度,进而验证了该方法的有效性.
Aiming at the continuous behavior recognition of human joint sequence, this paper proposes a method of human joint information behavior recognition and prediction based on BSCPs-RF (B-spline control points-random forest) .First, local linear regression and single-frame joint The normalization method was used to preprocess the joint sequence to eliminate the effect of jitter noise, displacement and scale. Then the B-spline control points were used as the speed-independent features of the joint sequence, and the real-time behavior sequence Improve the efficiency of sample collection.Finally, a method of behavior recognition and prediction based on random forest is adopted, and the multi-classifier combination is optimized by using integrated learning method to improve the recognition accuracy.Experimental analysis of the effect of different parameter values on the recognition effect is carried out in the test database MSR -Action3D and RGB-D equipment.The results show that the MSR-Action3D test results are better than some of the previous methods, while the real-time data test has high recognition accuracy, and then verify the effectiveness of the method Sex.