结合空间和运动特征的行为识别算法

Behavior recognition algorithm combining spatial and motion features

针对视频行为识别方法中运动特征不足导致识别精度低的问题,提出一种结合空间和运动特征的行为识别算法。通过空间和运动两路卷积提取特征。空间卷积采用Res2Net作为主干网络并添加注意力模块;运动卷积细粒度地分为局部和全局特征,构造自适应通道序列重构模块提取局部特征,采用逐帧特征信息作差后聚合的方法提取全局特征。将运动和空间特征信息融合后,经过全连接层输出类别。该模型在UCF101和HMDB51数据集上准确率达到95.6%和70.7%,与传统算法相比,识别精度得到一定提升,验证了该算法的有效性。

To solve the problem of low recognition accuracy caused by insufficient motion features in video behavior recognition methods,a behavior recognition algorithm combining space and motion features was proposed.Features were extracted by convolution of space and motion.Res2Net was used as the backbone network and the attention module was added in spatial convolution.The motion convolution was divided into local features and global features in fine granularity.An adaptive channel sequence reconstruction module was constructed to extract local features,and a frame-by-frame feature aggregation method was used to extract global features.After the motion and spatial feature information was fused,the categories were output through the fully connected layer.The accuracy of the model in UCF101 and HMDB51 data sets reaches 95.6%and 70.7%.Compared with the traditional algorithm,the recognition accuracy is improved to a certain extent,which verifies the effectiveness of the algorithm.