基于3D-SVD的时空行为定位算法

Spatio-Temporal Action Localization Algorithm Based on 3D-SVD

随着摄像头的普及,基于人工智能的行为分析技术在智能视频领域扮演着越来越重要的角色.现有的算法大多采用光流网络或者3D网络来获取行为的时间信息,但是光流网络和一般的3D网络计算量大,在同时进行分类和定位这两项任务时计算效率低.针对这一问题,本文构建了一个能够进行空间定位和分类的双流框架,在3D网络分支中采用SVD的思想分解3D卷积核以减少3D网络的参数,并利用动态规划算法高效的搜索最佳行为管道,在训练的过程中采用mixup算法对数据集进行扩充,增强训练的效果.最后,在UCF101-24和J-HMDB-21这两个被广泛使用的行为定位数据集上进行了实验验证,相比于基线算法,两个数据集的Frame-mAP分别提高了7.1%和4.8%,其中, J-HMDB-21在不同IOU下的Video-mAP分别提高了5.2%和4.8%.实验结果表明,本文提出的算法能有效提高行为定位能力,与其它算法相比获得了较好的结果.

With the popularity of video surveillance, action analysis technology based on artificial intelligence is playing an increasingly important role in the field of intelligent surveillance. Most of the existing algorithms depend on an optical flow network or a 3 D network to obtain the time information of actions. However, the optical flow network and the general 3 D network require a large amount of computation, and the computational efficiency is low when classification and localization are carried out simultaneously. To solve this problem, this study builds a dualflow framework capable of spatial localization and classification and follows the idea of SVD to decompose the 3 D convolution kernel in the branch of the 3 D network, thus reducing the 3 D network parameters. In addition, the dynamic programming algorithm is employed to efficiently search the optimal action tubes, and the mixup algorithm is used to expand the data sets during training, thereby enhancing the training results. Finally, experimental verification is carried out on UCF101-24 and J-HMDB-21, two widely used data sets for action localization. Compared with the baseline algorithm, the Frame-mAP of the two data sets is improved by 7.1% and 4.8%, and the Video-mAP of J-HMDB-21 under different IoUs is enhanced by5.2% and 4.8%. Experimental results show that the proposed algorithm can substantially improve the ability of action localization, with better results compared with other algorithms.