基于最大相关最小冗余的动作识别算法
An Action Recognition Algorithm Based on Maximum Correlation-Minimum Redundancy

为了提高动作的识别精度与鲁棒性，降低冗余特征，提高算法效率，设计了一种基于最大相关-最小冗余(Max-Correlation and Min-Redundancy，MCMR)的动作识别算法.首先，为了消除噪声影响，减少计算成本，利用符号聚集近似(SAX)技术将连续图像序列转换为离散符号；其次，为避免出现时间漂移问题，利用动态时间归整(Dynamic Time Warping，DTW)来计算符号特征的距离，提取符号序列的特征；然后，为了消除冗余的特征，定义了一个特征权重，根据权重对特征进行降序排列，引入最大相关-最小冗余技术消除相关性弱的特征，筛选出具有高相关性和低冗余的特征；最后，为了完成动作识别，根据筛选出的特征，利用k-近邻(K-Nearest Neighbor，KNN)进行分类器学习.结果表明：与当前动作识别算法相比，本文算法能够有效完成动作的识别与理解，具有较高的识别率，有效地降低了冗余特征，提高了算法的效率和鲁棒性.
In order to improve the accuracy and robustness of motion recognition, reduce redundant features and increase algorithm efficiency, an action recognition scheme based on MCMR (maximum correlation-minimum redundancy) is designed. Firstly, in order to eliminate the influence of noise and reduce computational cost, the continuous image sequence is transformed into discrete symbol representation by using the symbolic aggregation approximation technology. Next, to avoid the problem of time drift, dynamic time warping (DTW) is used to calculate symbol distance and to extract the character of a symbolic sequence. Then, in order to eliminate redundant features, a feature weight is defined, the features are arranged in a descending order based on weight, the maximum correlation-minimum redundancy is introduced to eliminate the weak correlation feature, and features with high correlation and low redundancy are selected. Finally, in order to perform action recognition, classifier learning is performed, using k-nearest neighbor based on the selected features. Test results show that compared with the current image action recognition, this algorithm can effectively accomplish action recognition and understanding. It has a high recognition rate, the redundant features are greatly reduced, and the efficiency and robustness of the algorithm are improved