Research And Implementation Of Driving Assistance System Based On Image Mosaic And Target Detection

In modern society,driving has been a part of people's daily life.The quantity of automobile in our country is about 200 million.With the number of the vehicle increasing rapidly,the traffic accidents have been common occurrence.Therefore,reducing the incidence of accidents has become people's fervent hope.In order to decrease the occurrence of driving accidents,reducing the difficulty of driving has become hotspot of research.At present,driver assistance system has owned by some of the vehicles,but the situation of the chassis has all been neglected,so it is more difficult to monitor the carbottom.If the camera is installed on the bottom of the car,there will be a shortcoming that the installation position is too low,resulting in a narrow field of view,and easy to be blocked when passing through obstacles.In this thesis,Image Stitching Algorithm is used to monitor the car bottom,which solves the problem that the common stitching algorithm can not satisfy the high consistency of the mosaic environment condition and the realtime performance of the algorithm.In this thesis and for the purpose of tackling this problem,a car driving auxiliary system based on image mosaic and target detection is designed,which also solves the problem of vision blind zone of car bottom.The driving auxiliary system of this thesis is divided into two modules,the chassis perspective module and the target detection module.The main contents of this thesis is composed of the following three parts:(1)Chassis perspective module.In the process of image initialization,adaptive gamma transform is introduced into the general stitching algorithm to improve the image quality,and the initialization module of adaptive detection threshold is used to ensure the effect of feature detection and description under high consistency conditions.Feature point position prediction module is designed and implemented in feature point matching,which utilizes the position relationship between feature points in the same frame and the continuity of displacement of feature points in adjacent frames to achieve fast and high accuracy matching of feature points.In homography matrix calculation,modified FSC algorithm is used to improve the speed and accuracy of matrix calculation,and homography matrix judgment module is introduced to ensure the robustness of stitching.(2)Target detection module.In this thesis,multi-task network is utilized to complete two tasks of target detection and semantic segmentation in one operation,and the network is also improved.In the front-end feature extraction module,a lightweight network is utilized to reduce the network running time greatly,while the back-end network extracts features again to make up for the shortcomings of the front-end feature extraction.At the expense of a small amount of accuracy,the network runtime is reduced by about 45%,which makes the algorithm more real-time.(3)Heterogeneous computing is used to speed up the operation in engineering implementation,and multi-threading is used to make full use of the performance of multicore processors.Finally,a real-time and accurate chassis perspective system is realized.Finally,the real-time operation of the chassis perspective module is realized,and the real-time performance of the network is improved in the target detection module,but it still needs real-time optimization to be applied in the actual system.