Research On Front Vehicle Detection Technology Based On Machine Vision

The number of private cars in China is rising year by year,with the rapid growth of car ownership,more and more cities are overcrowded,and traffic accidents occur frequently.Recently,automobile manufacturing company and Internet companies are committed to the research and development of advanced driver assistance system and intelligent driving,to protect people's life and property safety in the road environment as much as possible.On-road vehicle detection is an important research with regard to vehicle active safety and intelligent driving.Real-time and accurate vehicle detection plays a very important role in vehicle active safety such as: front collision warning,headway monitoring and automatic emergency braking system.In this thesis,a vehicle detection algorithm based on computer vision is proposed,algorithms can be divided into image preprocessing,hypothesis generation phase,hypothesis verification phase and distance measurement.(1)Image preprocessing: First,the effective area of the image is divided,and then the image after dividing the effective area is grayed.To reduce the noise in the image,three kinds of filtering algorithms are compared,and a median filtering algorithm with good denoising effect is selected.Then through the statistics of the image gray value,the image is classified according to the light conditions,and the histogram equalization processing is applied to the image of the over bright and dark scene to enhance the gray information.(2)Hypothesis generation phase: for the deficiencies of the biggest categorical variance method and the two largest interclass variance methods commonly used in segmenting vehicle shadows,an improved threshold segmentation algorithm is proposed.The algorithm combines a road surface grayscale sampling method and two largest interclass variance methods.First,the pavement sampling area is set,and the gray value of the collected road surface samples is counted.When the sample meets certain conditions,the threshold value is calculated using the sample data;otherwise,the road surface threshold is calculated using the two largest cluster-like variance method.After the binarization,there are many pseudo-vehicle areas,and some areas of pseudo-vehicles are screened using areas,rectangles,and quadrilateral measures.Then the two-valued image is hatched and merged.For the problem that adjacent vehicles are easy to miss in the process of merger,an improved algorithm of combination of shadow lines is proposed.(3)Hypothesis verification phase: By analyzing the advantages of the HOG feature and the LBP operator,a feature extraction method that fuses the HOG feature with the LBP feature and performs dimension reduction operation is proposed.It combines the advantages of the two operators and reduces the dimension of the feature.The extracted features are trained and classified using SVM classifiers.(4)Distance measurement: The related methods of front vehicle ranging are studied,including the imaging principle of the camera and the calibration of the camera.The external parameters and internal parameters obtained by the calibration of the camera are used to achieve the distance measurement of the vehicle in front of the vehicle under the static condition using the geometric model-based monocular distance measurement method.In this thesis,the proposed front vehicle detection algorithm based on machine vision is robust,not only for structured roads but also for unstructured roads,and the test results show that the algorithm has high detection rate and low false rate.