A Video Behavior Recognition Method Based on Deep Convolutional Features

Abstract

The invention discloses a video behavior recognition method based on depth convolution features, comprising the following steps: 1) extracting the dense trajectory of the video; 2) extracting the depth convolution space features of the video; 3) calculating the video optical flow and extracting the depth convolution 4) Perform space-time normalization and inter-channel normalization on the deep convolutional spatial features and deep convolutional temporal features respectively; 5) Normalize the spatial features and temporal features along the Time series pooling operation is performed on dense trajectories; 6) After pooling spatial features and temporal features are connected, LSTM network is used for classification. In the process of combining deep learning features and trajectory features, the method considers the timing information of trajectory features, can more effectively use video trajectory information, and make feature extraction more accurate. Finally, the LSTM network is used as a classifier, which advantageously improves the Accuracy of behavior recognition.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to a video behavior recognition method based on deep convolution features. Background technique [0002] Video, as a carrier that can carry more information than pictures, has gradually become one of the most important visual data in life. As a basic technology for video analysis and understanding, video behavior recognition technology is attracting more and more attention from scholars and engineers. On the one hand, behavior recognition is widely used in life and production, such as realizing intelligence and automatic driving. On the other hand, behavior recognition can promote the development of video analysis and understanding technology, and further promote the advancement of technologies such as network video transmission, storage, and network video personalized recommendation. [0003] Compared with image classification tasks, video-based classification needs to co...

AI Technical Summary

Problems solved by technology

Compared with manual traditional features, deep learning features can learn more distinguishing and hierarchical features, without considering trajectory features and their timing, ignoring motion characteristics, resulting in low accuracy of final classification

Images