Multi-view Perception For Unmanned Ground Vehicles

The research of computer vision perception technology in vehicle environment is an important research direction for unmanned ground vehicles.The research based on monocular vision is low cost,flexible and simple.But there exist the limitation of narrow vision field of road condition obtained by only one vehicle camera.It can get a wider range of vision using a rotating camera,but it cannot get the traffic information at the same time.Therefore,this paper combined with the specific application of unmanned ground vehicle,mainly focuses on how to get a better driving environment information perception using multiple cameras,including algorithms such as multiple cameras fusion in vehicles environment,the scheme of multiple vehicle cameras positioning and layout,binocular camera calibration and object distance perception when driving.The main work and research results are as follow:1.The multiple cameras fusion algorithm based on feature has a high requirement for feature extraction and feature matching.These algorithms mostly extract features in linear scale space,and the resulted image obtained after applying Gaussian filter can suffer from fuzzy boundaries and missing details.Therefore,this paper puts forward a multiple cameras fusion algorithm based on KAZE feature in nonlinear scale space.The algorithm improves KAZE feature in the nonlinear scale space combining the Gauge-SURF(G-SURF)feature description vector.It extracts KAZE feature points using nonlinear diffusion filter and AOS algorithm in video frames,generating 64-dimensional second-order multi-scale gauge partial derivatives feature,which describes vector in the gauge coordinate.We then stitch the images according to the feature matching transformation matrix,and generate a wide even panoramic scene.2.This paper also provides the scheme of multiple vehicle cameras positioning and layout from the principle of camera image-forming,camera parameters and the size of the car.This used to determine the optimal number and the location of cameras and get the best field of view in vehicles environment.It covers the largest area with the least amount of cameras.Then calibrating the binocular camera and applying the calibration parameters to the work of object distance perception in vehicles mobile environment.3.This paper conducts experiments in different datasets,and results show that the improved KAZE feature reduces the feature extraction time and feature matching rate has a certain improvement.The multiple cameras fusion algorithm can eliminate the shadow and distortion of the overlapping area in the process of fusion.The error of binocular camera object distance perception can be controlled in a certain range,it can be used in the unmanned vehicle driving environment,and provides a basis for changing the driving parameters to guide safe driving.