Research On Omnidirectional Vision SLAM Based On Vehicle-mounted Multi-camera System

In recent years,with the rapid increase of the use of vehicles in China,the loadbearing capacity of the road gradually tends to saturation,at the same time,it also brings many problems,such as traffic accidents,vehicle congestion,environmental pollution,etc.The current vehicle environment perception positioning mainly depends on Lidar and GPS sensors.Lidar is expensive and has limited sensing range,which makes it difficult to obtain omnidirectional environment information;GPS positioning accuracy is not high,and positioning information can’t be obtained under viaduct,tunnel and other scenes;however,the visual sensor has strong environmental expression ability,low cost and light weight,which has great potential to become the most important sensor of vehicle environment awareness and positioning.At present,the mainstream monocular and binocular vision sensor configuration scheme has many problems to be solved,such as narrow local imaging field of view,less information and not closely combined with vehicle motion state.Therefore,we propose an Omni-directional SLAM system for the perception of vehicle environment and the estimation of its own motion state.The Omni-directional SLAM system proposed in this paper consists of a forward binocular camera and three monocular cameras.The main research work and innovation are as follows:1.Key frame Selection Strategy: According to the fixed position of the vehicle camera and the motion of the camera frame,this paper proposes a strategy to insert the camera key frames in all directions.The initialization of the forward-looking binocular camera is the initial step of the whole system.2.Multi-Camera System and Vehicle Motion Estimation: In plane motion,the relative attitude of the vehicle between two positions can be expressed by yaw angle θ and polar coordinates (ρ,φ).The vehicle is modeled.According to its kinematic and geometric correspondence,the kinematic constraints of the vehicle are obtained.Combined with the rigid constraints between multi cameras,a multi camera system and vehicle motion estimation method is proposed.3.Multi-Camera Map Point-by-Point Cloud Registration: In this paper,a point set to point set ICP registration algorithm is adopted.The rigid constraint between multiple cameras is used as the initial value of the matching between the source point cloud and the target point cloud for coarse registration,and the optimal rotation translation matrix is solved by the least square iteration method for fine registration between the source point cloud and the target point cloud.Through experiments,the real-time accurate sparse point cloud map of vehicle surrounding environment is constructed.4.Local and Global BA Optimization: For Omni-directional SLAM problems with large amount of data,this paper proposes an optimization method combining local optimization with global optimization.The local BA optimization only includes the state of the latest frame and point in the sliding window.The global BA optimization runs in parallel with the local BA optimization at a lower frequency,and optimizes the camera frames which are removed from the sliding window but selected as the key frames in the global map,so as to prevent the drift of the whole system due to the cumulative error.