Research On Methods In Road Anomaly Detection And Avoidance For Automated Vehicles

From the research of interests and significance in the autonomous driving domain,it is shown that almost all the literature related to obstacle detection,recognition and avoidance are focused on the obstacles that appear on the roadways,such as buildings,vehicles and pedestrians.To this end,the dissertation focuses on road anomaly,i.e.,the unhealthiness of the road itself,which draws much less attention compared to the other obstacles and even get omitted by the academic researchers.Among the different types of road anomalies,pothole is the most annoying,intimidating and destructive one.Hence,this thesis is devoted to research on the methods in pothole detection and avoidance for self-driving cars or intelligent collected vehicles.According to the available image based methods for pothole detection,it is found that the detection approach is either based on image processing techniques followed by thresholding methods or machine learning algorithms,or based on deep learning approaches.This results in the relatively poor detection performance of the latter one compared to the latter pothole detection approach.However,the latter approach comes with the drawbacks of very complex models,large dataset required,and long training time.To take advantages of the two approaches,the vision based pothole intelligent detection algorithm is proposed,which relies on image processing techniques and deep learning models.Herein,the task of image processing applied is to extract ROIs(region of interests)from the original camera captures.And the deep learning model,i.e.,the learned Alex Net pothole detector,is used for classifying the ROIs.By leveraging the constructed testing dataset,the accuracies of the learned Alex Net pothole detector and proposed vision based pothole intelligent detection algorithm are 98.42% and 93.60%,respectively.In terms of road anomaly avoidance,the thesis takes two steps,i.e.,path planning and trajectory planning,by following the main solutions of obstacle avoidance.Unlike most of the path planning frameworks which rely on different kinds of sensors and are performed in the inertial frame,the dissertation aims to conduct the planning task directly in the image frame,and only relies on the partial imaging information captured by monocular camera.In the process of studying,two methods for distance estimation are proposed by using monocular camera.Moreover,a new lane lines detection method,and a new method based on ellipse for constructing cost map are also proposed.By incorporating those ingredients,the image based path planning algorithm is proposed.Finally,the two newly proposed distance estimation methods are validated through calibration,which result in an error of 3.32% and 6.99%,respectively.Also,the effectiveness of the proposed path planning algorithm is validated by using simulated traffic scene.It is known that trajectory planning is essential for a driverless car to follow an optimal path,which is generated by path planning.As there is only partial scene information available captured by a monocular camera,it is very difficult to estimate the location and velocity of the unmanned vehicle,and is prone to be lack of fidelity.To this end,a MPC based dynamic trajectory planning algorithm is proposed for the challenge.Compared to ordinary MPC based trajectory planning algorithms,the proposed one requires more computational resources,and consumes less of other kinds of sensors.By applying a discrete-time linear single track kinematic vehicle model,the setting of the hyper-parameters of the proposed MPC based dynamic trajectory planning algorithm is informed via simulating and analyzing that of the traditional MPC based trajectory planning,which facilitates the settings of the proposed algorithm when testing.To validate the effectiveness and the feasibility of the proposed road anomaly avoidance framework,road anomaly and a two-lane two-way road are simulated in lab environment,and two kinds of traffic scenes are considered with the existing of simulated road anomaly,i.e.,where the neighborhood lane is not occupied and where it is occupied by other vehicles.By virtue of the smart DIY robocar,it is successfully verified the effectiveness and feasibility of the proposed road anomaly avoidance framework.