Body orientation prediction method and device based on deep learning

Abstract

The invention provides a body orientation prediction method and device based on deep learning and relates to the technical field of video analysis. The method comprises the following steps: obtaining the position of each pedestrian in each frame in a frame sequence corresponding to a video to be detected; according to the position and a pre-established pedestrian space-time characteristic prediction model, extracting spatial features of each pedestrian through a convolutional neural network; according to the spatial features and the pedestrian space-time characteristic prediction model, extracting time characteristics of the frame sequence through a recurrent neural network of a door structure to obtain space-time characteristics of each pedestrian; and extracting direction characteristics in the space-time characteristics to obtain body orientation of each pedestrian. The method and device carry out modeling based on the spatial features and time characteristics of each pedestrian, and provide rich historical continuous change information to assist current-frame prediction, thereby improving body orientation prediction accuracy.

Description

technical field [0001] The present invention relates to the technical field of video analysis, in particular to a method and device for predicting the orientation of a human body based on deep learning. Background technique [0002] Intelligent video surveillance systems are widely used in the fields of urban security, traffic management, and environmental monitoring. The monitoring and analysis system for pedestrian behavior plays an important role in urban traffic management, special event prevention, and traffic safety. Among them, intelligent video surveillance is a multidisciplinary research direction such as signal acquisition and transmission, image processing, computer vision, machine learning and pattern recognition. At present, the research on the monitoring and analysis system of pedestrian behavior is very extensive, such as crowd density estimation, traffic statistics, pedestrian gesture recognition, special event detection and so on. However, pedestrian body o...

AI Technical Summary

Problems solved by technology

[0004] It can be seen that although the above-mentioned 2D method can complete the judgment of the orientation of pedestrians, but because only the information of a single frame is considered, this method has very low robustness to illumination and occlusion

The 3D method is difficult to be practical because it is unrealistic and unreasonable to obtain depth information and multi-angle shooting information in actual scenes.

Therefore, the judgment of human orientation based on the image features of a single frame loses a lot of useful information, making the prediction result less robust to illumination, occlusion, and multi-directionality, resulting in lower accuracy of human orientation prediction

Through the continuity of the video, the multi-frame time feature is used to judge the pedestrian's human body orientation. The multi-directional robustness of the pedestrian's forward movement is very low, which makes the accuracy of the human body orientation prediction relatively low, and cannot be applied to the actual monitoring system.

Images