摘要
There are some the issues caused by the diversity and complexity of anomalous events in video, a sequence recognition-based video anomaly detection method is proposed to better extract feature vectors and carve anomaly boundaries to improve the detection accuracy. In order to avoid annotating unusual clips or clips in the training video, which is very time consuming, the weakly labeled is invoked to train videos. The normal and abnormal videos as a whole sequence are used to accomplish the anomaly detection task. First, the frame rate and size of the video are unified, and the video is decomposed into RGB frames and optical flow frames. Next, the I3D model will be used as a feature extraction model to extract the feature vector of the video from the decomposed video frames. And then the feature vector is input into the Bi-LSTM model to learn the context informations between video clips, and the hidden layer states of the Bi-LSTM model is encoded as the feature of the video. Finally, the encoded feature vectors are input to SR-Net to obtain the video anomaly score, and the anomaly score is used to detect whether there are anomalous events in the video. Theoretical analysis and experimental results show that the proposed method achieves a video-level AUC detection accuracy of 85.5% and a false alarm rate of 0.8 on the UCF-Crime dataset. Compared with previous algorithms for anomalous event detection based on multi-instance learning, the detection algorithm in this paper has a higher accuracy and a lower false alarm rate. The proposed method provides better detection results in video anomalous event detection tasks.
学科与关键词
文献加载失败,请重试
重试