一种结合迁移学习和LSTM的视频暴力检测方法

A Video Violence Detection Method Combining Transfer Learning and LSTM

计算机视觉在安防领域发挥着重要作用,可以作为辅助手段对正在发生的暴力行为进行识别和报警,从而大大减少了视频监控对人的依赖。为了降低监控视频中暴力检测的运算成本并提高检测效率,在已有模型基础上提出一种利用迁移学习、长短期记忆网络(Long Short Term Memory,LSTM)来构建新的特征提取器,然后在新数据集上进行训练并微调部分参数,最终实现实时的视频暴力行为检测。提出的方法在RWF-2000数据集上取得了96.51%的准确率,优于传统方法,并且处理帧率能够基本满足实际需求。

Computer vision plays an important role in the field of security.It can be used as an auxiliary means to identify and alarm the ongoing violent acts,which greatly reduces the dependence of video surveillance on people.In order to reduce the operation cost and improve the detection efficiency of violence in surveillance video,a new feature extractor and time sequence modeling of video frame sequence were constructed by using transfer learning+Long Short-Term Memory(LSTM)method on the basis of the existing model,and then some parameters were trained and fine-adjusted on the new data set,and finally real-time video violence detection was realized.The proposed method achieves 96.51%accuracy on RWF-2000 dataset,which is better than the traditional manual feature method,and the processing frame rate can basically meet the actual demand.