3D Data Based Understanding Of Unstructured Scenes For Vehicle Nayigation

Scene perception is one of the most important tasks for unmanned vehicles’(UV) navigation. It is also one of the most challenge issues in artificial intelligence area.3D sensors are easy to use. They have high accuracies and provide abundant information.64beams lidar and stereo vision systems are getting popular in UVs’3D sensor systems.3D data is highly related to environment showing the property of randomness. Compared to structured scenes, there is no artificial mark in unstructured scenes. The complexity of3D object’s contour, bumpy road condition, disorder vegetation cause additional difficulties in scenes understanding. Therefore, studying3D data perception and procession are meaningful for expanding a UV exploring range and raising its capability of adaptation to complex scenes. Hence, this thesis investigates typical unstructured scene perception problems based on3D data. The following is main content and contributions:A64beams Lidar intrinsic parameters calibration algorithm and a Lidar extrinsic parameters self-calibration algorithm are proposed. By analyzing mathematical model of the Lidar emitters, we propose a variable distance compensation parameter which is linear to distance instead of the traditional constant distance compensation parameter. Using the variable distance compensation parameter makes intrinsic calibration more accurate. For extrinsic parameters calibration, we separate the common intricate calibration procedure into two steps:first, we take ground plane for calculate Lidar’s pitch、roll angle and transition in Z direction; second, we take advantage of pole-like object in a road and GPS records for calculate Lidar’s navigation angle and transitions in x-y direction. Combining the parameters, accurate extrinsic parameters can be achieved.In order to take a good understanding of rural scenes, a MRF (Markov Random Field) based road segmentation algorithm is proposed for UVs. Compared to road segmentation in a grid map, a graph based algorithm has longer detection range, higher accuracy and less expensive memory storage. Existing graph based algorithms just use single3D point features which are sensitive to noise and can not be implemented in bumpy rural areas. The proposed algorithm uses3D Lidar scan lines’ geometry features on xoy plane and Lidar’s data structure to separate every scan lines into small line segments. By analyzing line segments features, we build line segments’ potential functions and optimize the results by graph cuts. We test the proposed algorithm in real data sequence. Experiment results show that the proposed algorithm not only has higher accuracy but also has higher stability in a sequential detection. The proposed algorithm can be implemented in real-time.In order to taking perception and understanding of scatter vegetation wild scenes, we propose a multi-sensors fusion algorithm for obstacles classification. Firstly,3D data and2D data is fused for superpixel segmentation and depth upsampling. In iteration procedures,2D data and3D data guides each other to improve the performance of the algorithm. Secondly, we extract local features:including point clouds statistical, Lidar intensity, normalized vegetation differential index to comprise feature vector. A supervised learning is taken for training an SVM classifier. Experiments in real scenes show that multi-sensor based algorithm raises up the accuracy of the obstacles classification and improves the performance of the long range obstacles classification.In order to perception and understanding unknown complex terrain scenes, a algorithm framework is proposed. There are tasks of point clouds filtering, multi-frame combining, frames matching, point clouds recovery, point clouds rasterizing, traversability analyzing and obstacles clustering in the framework. In traversability analyzing, we take a UV size as calculation windows’size, computer the center grid property:if there are any two grids which their step height above the threshold and their gradient above the threshold too, then the center grid is labeled as a step obstacle. Taking all grids inside the window into computing slope and roughness and the traversability value.