Gait recognition method and device based on spatial-temporal feature fusion and storage medium

Abstract

The embodiment of the invention discloses a gait recognition method and device based on spatio-temporal feature fusion and a storage medium. The method comprises the following steps: acquiring a gait video sequence; extracting a pedestrian gait contour sequence from the gait video sequence; training the designed network model by adopting a pedestrian walking gait contour sequence; and performing forward reasoning on a to-be-detected pedestrian by adopting the trained network model to obtain gait features of the to-be-detected pedestrian. According to the method, the time sequence of the gait video sequence can be reserved, the parameter quantity of the network is further reduced, that is, the better effect is achieved on the aspects of the effect and the model complexity, in addition, the time and space characteristics of the gait sequence are fused, and the algorithm is more efficient. A gait recognition algorithm which is smaller in model and higher in recognition precision is realized.

Description

technical field [0001] The present invention relates to the technical field of gait recognition, more specifically to a gait recognition method, device and storage medium based on spatio-temporal feature fusion. Background technique [0002] As a unique biological characteristic of each person, gait has the characteristics of non-contact and identifiable at a distance compared with biological characteristics such as iris and fingerprints. Wide range of applications. [0003] At present, the research on gait recognition methods is mainly divided into two categories. The first category is based on a single image, which compresses a video sequence into a gait template image, such as a gait energy image (GEI). The method is simple and easy to train, but it lacks the time and fine spatial information of the video sequence; the second type is based on the video sequence method, which can usually retain the time and space information, but the model training difficulty and complexi...

AI Technical Summary

Problems solved by technology

[0003] At present, the research on gait recognition methods is mainly divided into two categories. The first category is based on a single image, which compresses a video sequence into a gait template image, such as a gait energy image (GEI). The method is simple and easy to train, but it lacks the time and fine spatial information of the video sequence; the second type is based on the video sequence method, which can usually retain the time and space information, but the difficulty and complexity of model training are higher than the first. The class method should be high, such as the existing 3DCNN model, which consists of 3D convolution and multiple fully connected layers

[0004] GaitSet is a method based on video sequences. It inputs the video series as a set, extracts features from each frame of images, and then uses the method of maximum pooling and mean pooling to fuse the features of each frame. Although this makes the network more It is simple and flexible, but it also lacks the time sequence information of the gait video frame, and this information is very important for the improvement of the gait recognition effect

Images