A method for human behavior recognition based on 3D deep convolutional network

Abstract

The invention belongs to the field of computer vision video action recognition, and discloses a method for human action recognition based on a 3D deep convolution network. The method first divides a video into a series of continuous video segments; then, divides the continuous video segments Input to the 3D neural network composed of convolutional computing layer and spatio-temporal pyramid pooling layer to obtain continuous video segment features; then calculate the global video features as behavior patterns through the long and short memory model. The technology of the present invention has obvious advantages. By improving the standard 3-dimensional convolutional network C3D and introducing multi-level pooling, feature extraction can be performed on video clips of any resolution and duration; at the same time, the robustness of the model to large behavior changes is improved, It is beneficial to increase the scale of video training data while maintaining video quality; embedding correlation information through each motion sub-state improves the integrity of behavior information.

Description

technical field [0001] The invention belongs to the field of computer vision video recognition, in particular to a method for human behavior recognition based on a 3D deep convolutional network. Background technique [0002] In the field of computer vision, the research on action recognition has gone through more than 10 years. As an important part of pattern recognition, feature engineering has always been dominant in the field of behavior recognition. Before deep learning, scientists Evan Laptev and Cordelia Schmid of the French computer vision institution Inria made the most outstanding contributions to the learning of behavioral features. Similar to the ILSVRC Image Recognition Challenge, the action recognition-based challenge THUMOS continues to refresh recognition records every year. The behavioral feature calculation method introduced by Inria has always been among the best. Especially in 2013, Dr. WangHeng of Inria proposed a trajectory-based behavior feature calc...

AI Technical Summary

Problems solved by technology

[0003] To sum up, the problems existing in the existing technology are: the existing 3-dimensional convolutional network exists: the network can only extract the sub-motion state; every small segment of the video belongs to the same behavior category; the existing behavior recognition network Only the sub-motion state can be extracted; every small segment of the video belongs to the same behavior category; the scale and duration of each input video segment must be fixed. processing; at the same time, the network learns short-term motion features, lacking complete behavioral information

Images