Far-infrared pedestrian training method based on self-similarity gradient orientation histogram

Abstract

The invention discloses a far infrared pedestrian training method based on a self-similarity gradient orientation histogram. The invention helps resolve the problem that for HOG (Histogram of Orientation Gradient) characteristics, mutual relations between local blocks of images cannot be described, thereby bringing poor representation capability of characteristics. The invention provides a new self-similarity measurement HOG (Self-Similarity HOG, SHOG) feature in order to enhance the characterization capability of the HOG feature. Noise samples are removed based on SHOG feature clustering, and high-quality samples are obtained. Further, linear support vector machine models of four branches are acquired based on the SHOG features. The method comprises a feature improvement module for improving the current HOG features by increasing similarity measurement between image blocks, a feature extraction and noise sample rejection module for rejecting noise samples based on SHOG feature clustering, and a four-branch linear support vector machine training module based on SHOG features. The method can give consideration to both the detection and classification accuracy and the classification speed.

Description

technical field [0001] The invention belongs to the fields of computer vision and pattern recognition, image processing and computer vision, and in particular relates to a far-infrared pedestrian training method of self-similarity gradient orientation histogram. Background technique [0002] Because road pedestrians naturally lack effective shelters, in various types of traffic accidents, collisions between road pedestrians and moving road vehicles have particularly serious consequences. Especially in the situation of insufficient light at night, this type of collision occurs from time to time. Since the far-infrared camera does not depend on the light source for imaging, it can be imaged simply according to the temperature difference. Therefore, using the vehicle-mounted far-infrared camera to capture the pedestrian target in front of the vehicle, develop a vehicle-mounted pedestrian detection system based on image processing and pattern recognition technology for assisted ...

AI Technical Summary

Problems solved by technology

[0009] The purpose of the embodiments of the present invention is to provide a far-infrared pedestrian training method based on the self-similarity gradient oriented histogram, aiming to solve the unsatisfactory recognition accuracy and real-time performance of the existing far-infrared camera-based vehicle-mounted pedestrian classification method. It is difficult to balance the problem of robustness

Images