A Recognition Method of Climbing Behavior Based on Convolutional Neural Network

Abstract

The invention discloses a convolutional neural network-based climbing behavior recognition method, which is applied to the field of target recognition, and aims at the problem of low detection accuracy in the prior art in the behavior recognition of pedestrians climbing over railings; the invention uses drawing and character size The same bounding box overcomes the disadvantages of low real-time performance and immutable size of the bounding box in the traditional target detection method; the Yolo target detection network is used for image feature category prediction, and the GOTURN network is used for target tracking; finally, through prior knowledge The method quickly uses the relative positional relationship between the railing and the track point set to determine whether it is a crossing behavior, and if it is a crossing behavior, output the crossing label and issue a warning.

Description

technical field [0001] The invention belongs to the field of target detection, in particular to a behavior recognition technology. Background technique [0002] For multi-type target scenes, the target detection method aims to accurately distinguish the category and position of the target in the image, and the two-stage method can solve this kind of problem. Researchers mainly generate candidate frames through the Region Proposal method, and then perform coordinate regression prediction based on the candidate frames. Ross Girshick and others used CNN network to extract image features, from the experience-driven artificial feature paradigm HOG, SIFT to the data-driven representation learning paradigm, improving the ability of features to represent samples, and using supervised pre-training under large samples and fine-tuning with small samples The way to solve small samples is difficult to train or even overfitting, which improves the accuracy of target detection to a certai...

AI Technical Summary

Problems solved by technology

The generative method uses a feature model to describe the appearance characteristics of the target, and then minimizes the reconstruction error between the tracking target and the candidate target to confirm the target; the generative method focuses on the feature extraction of the target itself, ignores the background information of the target, and uses Target drift or target loss is prone to occur when drastic changes or occlusions occur

Images