Research On Key Technologies Of Vehicle-based Environment Sensing System Based On Vision

With the development of the society,cars have become an important way of travel.However,while the car brings convenience to people,it has caused a number of traffic problems,so the vehicle is driven by advanced assisted driving technology.At present,more accidents are caused by the driver's inability to grasp the external environment of the vehicle in time.The existing laser radar environment-aware sensors are difficult to popularize due to high price,and the millimeter-wave radar environment-aware sensor has a very limited detection range due to wavelength.The detection accuracy of the target is low and the target type cannot be identified.The ultrasonic radar environment sensing sensor has a slow response speed,low precision,small detection distance and cannot meet the requirements of the vehicle environment perception,and the vehicle camera sensing environment has rich information and low cost.The advantages of mass production play an important role in vehicle environment perception.In order to improve the ability of the vehicle environment to sense,vision-based environment-aware technology has been widely studied,among which the key technologies are vehicle detection technology,vehicle identification technology,traffic sign detection technology and so on.Analogous to the movement decision-making situation in which human beings are in an unfamiliar environment,it is first necessary to clarify their specific location information,secondly to perceive the obstacle information around them,and finally to specify the specific information of the obstacles on the route of the path.Based on this,the vision-based environment-aware technology mainly involves real-time acquisition of its own viewing information,thereby reducing the collision probability of the vehicle with surrounding obstacles during the driving process,real-time positioning of the vehicle,thereby clarifying the driving direction and obtaining the information of the preceding vehicle,thereby assisting Action decision.The main work and research contents are as follows:(1)Aiming at the perception of the vehicle's own viewing information,the traditional car panoramic image stitching method has the problems of cumbersome parameter calibration and unfavorable real-time calibration.This paper proposes a key frame-based registration algorithm,which mainly includes the following aspects: 1)According to the vehicle The relative position between cameras is fixed for a period of time.It is proposed to extract feature point pairs based on continuous key frames to expand the quality and quantity of matching pairs,so as to compensate for the low image resolution and the sparseness of scene feature points.2)For the problem ofmismatching,a combination of weight and filtering is proposed to repair the matching pair.Firstly,according to the case where there is one-to-many matching in the matching pair,the reverse matching point is eliminated,and then the improvement scheme of the RANSAC algorithm is proposed,so that the erroneous matching pair and the quality matching pair are well performed.The search is performed;3)Finally,the inter-image registration parameters are solved in a large feature point pair set.(2)Aiming at the problem that the vehicle monocular visual pose estimation has poor adaptability to dynamic scenes,a vehicle visual pose estimation algorithm based on dynamic obstacle filtering is proposed.The method trains a model for detecting the vehicle in front in a machine learning manner and combines the prior knowledge to distinguish whether the detected vehicle is a dynamic background,and finally removes the dynamic background from the extracted feature points.The region and the dynamic adjustment of the threshold of the extracted feature points according to the size of the dynamic obstacle effectively avoid the disadvantages caused by the detection module,thereby ensuring that the total number of extracted feature points is not greatly reduced.(3)The overall process of license plate recognition for sports vehicles is designed by ourselves.It is difficult to operate the license plate for sports vehicles.The Haar+AdaBoost cascade classifier method is used to detect the license plate information.In combination with the subsequent character recognition,the license plate image is too fuzzy to be correctly identified.The problem of license plate characters is used to traverse the license plate position by means of pre-divided detection area to improve real-time detection performance;in the character recognition module,for the distantly blurred characters,the deep learning model is used to enhance the fuzzy characters by means of data enhancement.The ability to identify,then propose improved methods for LeNet-5 deficiency,and introduce the corresponding technology to make the model better adapt to the classification task of license plate character recognition.