Research On Key Problems In Localization And Mapping Technology For Autonomous Indoor Parking

As an important part of automatic driving system,an autonomous parking system can effectively solve the“last mile”problem drivers are faced with.Normally,an autonomous parking system includes modules such as environment perception module,decision-making module and vehicle-control module.Among them,the environment perception module utilizes various sensor devices to determine the location of the autonomous vehicle and perceive its surrounding environment,which is the basis of the subsequent decision-making and vehicle control modules.According to the different parking scenarios,the environment perception module has different implementation approaches.In outdoor scenarios,due to the accessibility of the high-precision global navigation satellite system,the position of an autonomous vehicle is relatively easy to achieve.However,in the indoor scenario,due to the obstruction of the wall,the autonomous parking system cannot achieve this goal based on the satellite signal.Therefore,in an indoor parking environment,autonomous vehicles rely on the on-board sensors or indoor wireless sensors.Compared with indoor wireless sensors,the on-board sensors-based schemes are of high flexibility.With the SLAM（Simultaneous Localization And Mapping）technology,the vehicle can localize itself successfully in the unknown indoor parking environment and construct its surrounding environment.In view of the performance complementarity of camera and IMU,the VI-SLAM（visual inertial SLAM）system constructed by fusing the data collected by them has become the mainstream.However,traditional SLAM systems mostly rely on low-level visual features in the scene or semantic objects in front of the vehicle（such as pedestrians,vehicles,etc.）.When deployed in a complex parking environment,such SLAM systems are prone to problems such as low localization accuracy or even tracking failure.Meanwhile,the built map lacks the most important ground objects such as parking-slots.Therefore,these SLAM systems are not suitable for indoor autonomous parking.Based on the above analysis,this dissertation addresses key issues to be solved in the SLAM systems suitable for autonomous indoor parking,that is,how to accurately locate the autonomous vehicle and build a high-precision semantic map of its surrounding environment by using only multi-source on-board sensors in the indoor parking environment with weak GNSS signal.The main contributions of this dissertation are summarized as follows:（1）In order to effectively detect the semantic objects on the ground,the fisheye images taken by the surround-view system should be merged to generate a seamless bird’s-eye surround-view image.This dissertation proposes an extrinsic calibration scheme with high calibration accuracy and low requirement for calibration environment.This scheme can complete the calibration of surround-view systems in the geometric domain so as to generate high-quality images.The experimental results show that the proposed calibration scheme of the surround-view system is effective.（2）This dissertation designs a VI-SLAM system:VIS_SLAM,for autonomous indoor parking using three kinds of sensor data:low-level visual features extracted on front-view images,motion data output by IMU and parking-slots detected on the surround-view images.The SLAM system estimates the vehicle trajectory,and constructs the three-dimensional map with ground objects of the surrounding environment based on the localization results.The experimental results show that in the condition of limited computing resources,VIS_SLAM system can ensure high localization accuracy and output a high-precision semantic map with parking-slots.（3）By fusing various ground objects,this dissertation designs a robust SLAM system:MOFIS_SLAM,for complex indoor parking scenes.As far as we know,MOFIS_SLAM system is the first SLAM system for autonomous indoor parking that can integrate all kinds of surround-view semantic features in a tightly-coupled optimization framework.In addition,in MOFIS_SLAM system,the surround-view data association method is based on spatial location and feature matching,which can obtain robust data association results.The experimental results show that compared with VIS_SLAM system,MOFIS_SLAM system has higher localization and mapping accuracy in indoor parking environment.（4）In order to objectively evaluate the performance of different SLAM systems for indoor parking environment,this dissertation establishes a large-scale benchmark dataset named Be VIS.In addition to the multi-source sensor data,the dataset also provides the groundtruth of the vehicle trajectory,which provides a new benchmark for objectively and quantitatively evaluating the localization accuracy of SLAM systems.In addition,this dissertation validates the effectiveness of our proposed groundtruth trajectory acquisition method by measuring the reprojection error and pose fluctuation of the vehicle.The experimental results show that MOFIS_SLAM system outweighs other competitors in localization and mapping accuracy in Be VIS dataset.The publicly available dataset Be VIS can promote the relevant research of SLAM systems for indoor parking environment.