Research on Human Abnormal Behavior Detection Based on Deep Learning

In order to make full use of the effective information in the video and improve the recognition rate of abnormal human behavior in complex scenes, we use a mixed Gaussian model to detect clear foreground moving target contours and perform Gaussian filtering on them to remove the effects of noise in the scene . By calculating the center point of the foreground pixel, and drawing a bounding box based on this, the key area of human motion in the video is extracted. Then we use the Farneback dense optical flow algorithm to obtain spatiotemporal information. By combining CNN and LSTM, a CNN-LSTM hybrid two-stream network model based on the Dropout mechanism is established., input the original image and the superimposed optical flow image of the key area of the video sequence motion to learn the dynamic and static features and timing information in the spatiotemporal information. The weighted fusion method is used to perform weighted calculation on the Softmax output of the two-way network to obtain results. The tresults show that the accuracy of the behavior classification reached 91.2%, and the recognition rate of abnormal behavior was 92%. Compared with the three models in the article, the improvement was 6% 8.3%, 3.4%.