A Human Object Detection Method Based on Regional Fully Convolutional Neural Network

Abstract

The invention discloses a human body target detection method based on a regional full convolutional neural network, which calibrates the upper body of the human body target, that is, the position of the human head and shoulders model, as a calibration frame, thereby reducing the probability of the human body target being blocked and reducing the missed detection rate; At the same time, five different image scales are set for conversion, and two different aspect ratios of {0.8, 1.2} and five different scales of {48, 96, 144, 192, 240} are selected as the rules for generating anchors, which further reduces the leakage. inspection rate. In addition, the present invention also calculates the loss value of candidate frames in each region of a human target image, and selects the candidate frame in region B before the largest loss value as a difficult sample, and its loss value is fed back to the regional full convolutional neural network model, using random The gradient descent method updates the parameters of the regional fully convolutional neural network model to improve the accuracy of human target detection in complex scenes and reduce the missed detection rate and false detection rate.

Description

technical field [0001] The invention belongs to the technical fields of computer vision, pattern recognition, machine learning, etc., and more specifically relates to a human target detection method based on a regional fully convolutional neural network in a monitoring scene. Background technique [0002] In recent years, with the advancement of technology, various industries have begun to pay more and more attention to security issues. In important areas such as banks, airports, subways, stations, communities, and public places, people have installed surveillance cameras for video surveillance. These monitoring cameras are generally installed in a higher position to monitor from a bird's-eye view. The monitoring scene we are talking about refers to the monitoring picture taken in this situation. [0003] In general, people are the main body of surveillance scenarios, and the tracking of human objects and subsequent behavior recognition and analysis are heavily dependent o...

AI Technical Summary

Problems solved by technology

[0007] Although the R-FCN method has achieved good detection results in general target detection and human target detection, there are still some problems. For example, when there is occlusion between human targets, there is a problem of detecting two people as a single person. situation, resulting in missed detection, and when the human target scale is small, there are detection failures such as missed detection

In addition, for human targets in some complex monitoring scenarios, such as monitoring scenarios with complex backgrounds, large numbers of human targets, and serious human occlusion, existing human target detection methods still have a certain degree of missed detection and false detection.

Images