Vehicle Detection Method Research Based On Multi-feature Fusion Under Complicated Environment

Intelligent Transportation Systems (ITS), which aims at enhancing the safety andefficiency of road transport system, applies a variety of techniques like image processing,artificial intelligence, the embedded, sensor and pattern recognition to solve problems ofheavy traffic and traffic accidents. As an important part of ITS, moving vehicle detection hasattracted lots of focuses so far. During the detecting processes, they may have to deal withmany different problems, such as the influences from background environment, complicatedlight and mutable weather and the weak robustness of the methods. Therefore, it is required todetect the characteristics vehicle possessed, the road, the weather and other features toimprove the robustness of vehicle detection.Concentrated on the poor real-time and robustness for vehicle detection, this thesisproposes the vehicle detection method based on multi-feature fusion under complexenvironment which considered about variety effect factors. Obtained by the experiments, theproposed methods have achieved good performances. The main works in this paper aredescribed as follows:(1) Propose a segmentation method based on shadow feature of interested area. Firstly,confirm and separate out the driving fields. Secondly, make histogram estimate on gradationvalues of the road and carry out the threshold of shadow of the vehicle bottom. Then conductedge extraction on the shadow of the vehicle bottom through the extraction algorithm basedon the rate of change. At last, establish the vehicle interested region by the vehicle shadow.(2) Propose a vehicle detection method based on the feature fusion of edge contour,corner and texture. To cope with the light impact under complex circumstance, applyhistogram extension as pre-processing to be the basis. Employ K-R corner extraction to pickup corner features. Use improved fractal dimension calculation method to extract out thetexture features. Apply wavelet coefficients decomposed by morphological Haar wavelet toestimate the contour features. Then according to the eigenvalue of statistical samples, buildvehicle judgment formula fused these three features above through the various feature weightscalculated by Mahalanobis distance method. Therefore, this proposition mainly focuses on theextraction of the three eigenvalues of interested region and the vehicle judgment formulaaiming at deciding whether there are vehicles existed in the region.(3) Propose a vehicle detection method based on lights at night. Exclude parts of effectsfrom interfering light resource through lights’ gray features and the priori knowledgecharacteristics of light. Construct the object chains of candidate headlights. Conduct lights’ match based on the method of Euler distance and the priori knowledge characteristics of lightspair. Finally, establish region for the successfully matched vehicles.