Research On Image Based Vehicle Feature Extraction And RBF Network Recognition

With the rapid development of economy,the output of cars increases year by year.While bringing convenience to people,it also brings great pressure to road traffic,resulting in frequent traffic accidents.Therefore,the research on ADAS has become the focus of people's attention.The system represented by ADAS applies a variety of sensors in hardware,such as ultrasonic sensors,visual sensors,radar,GPS,etc.During the driving process,the system will perceive the vehicle's own state and detect the vehicle ahead and surrounding environment to reduce vehicle offset and avoid collision and other accidents.The vehicle environment perception system is the key of the vehicle safety assistant driving system,among which target vehicle recognition is one of the most important components of the environment perception system.Obtaining accurate and effective target vehicle information in front of the road can provide strong technical support for the safe driving of vehicles.The radar system used in traditional vehicle identification has strong adaptability to the environment,but it has the disadvantage of high noise of original data and high error detection rate.Therefore,the vision-based vehicle identification method is adopted,which is rich in information,low in cost and has great advantages in object detection and identification.This thesis mainly studies image-based vehicle detection and recognition.The specific research work is as follows:(1)process the road images collected by the CCD camera to realize the preliminary processing of vehicle detection based on the combination of vehicle taillights and shadows proposed in this thesis.It mainly includes five image pre-processing steps: determination of ROI,image graying,color image segmentation,morphological processing and filtering processing,extraction of taillight pairs and vehicle bottom shadow features,and elimination of noise points and other interference information.(2)a method based on vehicle taillight segmentation and vehicle bottom shadow is proposed to extract vehicle features,so as to determine the possible areas of vehicles.The taillight pair is extracted from the image to determine the width information of the vehicle,and the height information of the vehicle is further obtained by combining the structure size relation of the vehicle.According to the characteristics of vehicle bottom shadow,the threshold segmentation method was used to mark the vehicle bottom edge.In addition,for different road environments,a decision function is constructed and different weight coefficients are set up to ensure that the hybrid feature extraction method combining taillight and shadow can be used in different situations,providing an effective area for the accurate recognition of subsequent vehicles.(3)RBF neural network was constructed to identify vehicles in the identified effective areas.The regional description and shape characteristic parameters reflecting the vehicle characteristics were extracted,and the RBF neural network recognizer was designed with 19 characteristic parameters as input variables,such as cosine transform descriptor,independent invariant moment,regional eccentricity and ratio of regional length and axis,to realize the effective recognition of vehicles in the region of interest.The positive and negative sample base of vehicle image was established,the RBF neural network recognizer was trained and the algorithm was convergent,the accuracy rate was up to 94%.The neural network vehicle recognizer is verified by random test samples,and the forward detection rate and error curve show that the proposed research method can reliably identify vehicles in the effective region.