双流网络构架行为识别隐含层模型仿真

Double Stream Network Architecture Behavior Recognition Hidden Layer Model Simulation

复杂场景中特征的有效提取一直是行为识别的难点,动作和行为的精确表达也是影响识别结果的重要因素。针对当前行为识别模型只能选择单一特征进行行为识别,对动作和行为的表达有误,导致识别精确度、召回率低、识别速度慢等问题,提出双层网络架构和隐含层相融合的行为识别模型。采用CNN和RNN构建双流网络架构,用于抽取视频动作片段的外观特征和时间特征,同时在双流网络架构中添加一个隐含层,以便更有效地对特征进行描述。依据求和的形式融合运动特征和外观特征,构建运动行为组合特征,将运动行为组合特征输入到支持向量机分类器中来完成行为识别,在UCF101、UCF50数据集上进行行为识别实验。实验结果表明,所提模型有效提高了行为识别率和召回率,识别速度也优于对比模型。

At present, the behavior recognition model can only select single feature to recognize the behavior, resulting in low recognition accuracy, low recall rate and slow recognition. This paper focuses on the behavior recognition model integration combining the dual-layer network architecture with the hidden layer. At first, CNN and RNN were used to construct the dual-flow network architecture which was applied to the extraction of the appearance characteristic and time characteristic of motion segment of video. Meanwhile, a hidden layer was added to the dual-flow network architecture, so as to describe feature more effectively. According to the form of summation, the motion feature and appearance feature were integrated to construct the compound feature of motion behavior. Finally, the compound feature of motion behavior was input into the classifier of support vector machine to complete the behavior recognition. Thus, the behavior recognition experiment was performed on UCF101 data set and UCF50 data set. Simulation results show that the proposed model effectively improves the behavior recognition rate and recall rate. Meanwhile, the recognition speed is better than the comparison model.