Subregion Multifeature Fusion Oblique Vehicle Detection Algorithm

Abstract

Reducing traffic accidents and relieving traffic congestion are two main topics for Chinese road traffic management. Vehicle vision-based vehicle detection and tracking technology is the primary link in road environment perception. The detection technology for forward vehicles is becoming increasingly mature, and commercial products have appeared. Detection is more complicated for oblique vehicles. This research mainly studies the detection and tracking of oblique vehicles. The main work is as follows: first, the image preprocessing method in image processing technology is mainly studied. With focus on research of lane line detection algorithm, an improved lane line detection algorithm combining Hough transform and K-means clustering is proposed to improve the accuracy. With the lane line taken as the standard, the vehicle detection area is divided and combined with the ROI area extraction to further narrow the detection range. With the aim of performing vehicle detection, a two-level optimization algorithm is proposed. First, the method of initial inspection and machine learning detection verification of feature detection is presented. The vehicle target feature and image vertical edge feature fusion are used to generate the suspected target region. An improved dual threshold segmentation algorithm is then proposed to improve the effect of shadow feature extraction. The cascading AdaBoost classifier is combined to generate the target verification area. To address the real-time problem, kernel principle component analysis using Gaussian kernel function is applied to reduce the dimensionality of image feature and improve real-time performance. Experimental results show that the identification accuracy is up to 90% for oblique vehicles when the optimization algorithm of the mixing of shadow feature and image vertical edge feature is used. In good weather, the accuracy is up to 95% when using the AdaBoost classifier combined with HOG + Haar-like feature; the detection time is reduced by 26% when kernel principal component analysis is used. The result of this project is valuable for dynamic oblique vehicle detection.