LASRC-ODP降维算法在行为识别中的应用

Application of LASRC-ODP dimension reduction algorithm in action recognition

针对分类识别算法在行为识别中存在识别率低和实时性差的问题,提出了一种线性近似稀疏表示分类的正交鉴别投影(LASRC-ODP)算法用于行为识别。LASRC-ODP算法将高维数据投影到低维空间时,最小化类内残差和最大化类间残差,同时利用投影矩阵的正交约束来增强鉴别结果;与LASRC分类相结合,将训练样本构成过完备字典,利用L_2范数求解稀疏系数,优化了求解复杂度、加快了计算速度,得到特征更易区分的样本、最后根据L_1范数和残差找出对应类别,保证了强鲁棒性。采用KTH行为数据库进行实验,可使LASRC分类时识别率为97. 1%,实验结果表明该算法识别率高、抗噪和鲁棒性强,为行为识别的研究提供了一种新思路。

Aiming at the problem that classification algorithm has low recognition rate and poor real-time performance in action recognition,this paper proposed a novel method based on an orthogonal recognition projection algorithm with linear approximation sparse representation(LASRC-ODP).The LASRC-ODP algorithm minimized with-class residuals and maximized between-class residuals when projecting high-dimensional data into low-dimensional space,while improving the discriminant results by using orthogonal constraints of the projection matrix.And then it combined with LASRC classification making the training samples a complete dictionary,and by resolving to the sparse coefficients with L 2-norm,which could optimize the solution complexity,speeded up the calculation speed,and got the sample with characteristics more easily distinguished.Finally,it found the corresponding categories according to L 1-norm and residual.In the KTH behavior database,the proposed algorithm can make the recognition rate of LASRC up to 97.1%,and the anti-noise ability is strong,which can be used for the recognition of action.