Research On Unstructured Road Obstacle Detection Based On The Trafficability Of Unmanned Ground Vehicle

As the crystallization of modern high technology,unmanned ground vehicle has the ability of environment perception,decision planning,autonomous control,etc.It can move autonomously in complex environment,and is widely used in civil and military scenarios such as material distribution,terrain exploration,patrol and sentry,ammunition transportation,etc.It has great research and application value.As an important part of the unmanned ground vehicle,obstacle perception plays the role of the "eyes" of the unmanned platform,and is the primary factor for the unmanned ground vehicle to achieve safe driving.This paper combines a school-enterprise cooperation project,through the analysis of the ground unmanned platform operation scenario,the typical obstacles that have an impact on the safety of off-road vehicle driving as the research object,mainly for the generalized typical obstacles and some other general obstacles to carry out the research of obstacle detection and barrier crossing passability,the main research content is as follows:Firstly,the scenarios encountered by unmanned ground vehicle on unstructured roads in the wild are analyzed,and two typical obstacles,positive and negative,which have a serious impact on vehicle trafficability,are selected as the research objects.Considering that the obstacle surmounting ability of the 6x6 wheeled unmanned ground vehicle is relatively strong,theoretical research and analysis are conducted on its obstacle surmounting ability.For positive obstacles,the relationship between obstacle crossing height,wheel radius,and adhesion coefficient is established;For negative obstacles,the relationship between obstacle crossing width and wheelbase and wheel size is established.Secondly,the study of multi-frame laser point cloud-based obstacle detection is launched.Considering that unstructured roads do not have a flatter surface like structured roads and will have certain unevenness,some point cloud processing algorithms under structured roads are not highly applicable to non-scenes.In terms of ground point and obstacle point separation,the model-based ground fitting commonly used for structured roads is abandoned,and this paper incorporates the advantages of multi-scale rectangular raster division under Cartesian coordinate system and sector raster division under polar coordinate system to achieve obstacle detection based on raster height difference and minimum height.In order to compensate for the lack of detail caused by the sparse point clouds,the point cloud density is enhanced by implementing a before-and-after frame association based on IMU.Thirdly,in visual detection,convolutional neural network is used to achieve target detection.Considering the scenes of unstructured roads,where the size and aspect ratio of obstacles vary greatly,a one-stage Yolo X network without pre-scanning frames is used,and three aspects of the backbone network,neck network,and loss function are improved from the perspective of feature extraction and bounding box regression to effectively improve the accuracy of visual obstacle detection.Fourthly,the sensor fusion scheme is explored,and through the trade-off of various fusion schemes,a simple,effective and practical target-level fusion method is adopted to obtain more accurate obstacle information by using the property of laser point cloud with its own accurate position coordinates to compensate for the lack of obstacle depth and distance information in visual detection.At the same time,in order to compensate for the problem of different estimation of obstacle scales due to observation angles,a target tracking method based on Deep SORT is implemented on the basis of target detection,and a fusion result judgment strategy is designed to determine whether obstacles can be crossed.Finally,an unmanned ground vehicle model is built in ADAMS.Based on the judgment of obstacles by fusion detection,a corresponding scenario is established in ADAMS,and obstacle clearance simulation verification is conducted to verify the correctness of the fusion detection results,and the theoretical obstacle clearance value of the unmanned platform is verified.This paper takes the unmanned ground vehicle as the research object,studies the mechanism of its crossing positive and negative obstacles,and implements the fusion detection of positive and negative obstacles using laser radar and camera.When detecting obstacles,it not only avoids obstacles,but also provides a reference for whether obstacles can be crossed.It fully utilizes the off-road performance of the ground unmanned platform,providing a new idea for the obstacle detection research of unmanned platforms,Finally,the rationality of obstacle surmounting mechanism research and obstacle surmounting judgment is verified through ADAMS obstacle surmounting simulation.