Research On Target Detection And Tracking Algorithm Based On Lidar For Autonomous Driving

In recent years,with the rise of artificial intelligence and the continuous development of robot technology,driverless technology has become a global research hotspot.Unmanned driving technology mainly uses cameras,lidar,positioning inertial navigation system(IMU)and other sensors to realize real-time perception of the surrounding environment.Target detection and tracking is an important part of autonomous vehicle environmental awareness,and also an important basis for autonomous vehicle path planning and decision-making.At present,the research of obstacle detection and tracking mainly focuses on computer vision and lidar.Although the vision based deep learning target detection method has been widely developed in recent years,but this kind of method is easily affected by the light,which will lead to poor detection and tracking effect in the case of weak light.Lidar,with its special ability to obtain the basic shape,distance measurement and location of obstacles,has been recognized consistently in the detection and tracking of autonomous vehicles.In this paper,velodyne 32 line lidar is used as the data source to study the target detection and tracking algorithm applied to unmanned driving.The main work is as follows:(1)Point cloud scene has a large amount of data,so it is very time-consuming to process the data flow of point cloud in real time.In order to reduce the number of points and roughly remove background points,it is an important step to recognize scanning points as surface or non surface objects.In this paper,a slope based multi threshold detection algorithm(Multi Threshold Detection Based On Slope,MTD-BOS)is proposed.The proposed multi threshold algorithm is a method based on the slope of two points of the scan line,and uses different thresholds for different distance points to avoid false or missed detection of ground points.(2)Due to the working characteristics of lidar,the scanned points are often sparse and the degree of sparsity will increase with the increase of distance,which makes the point cloud data difficult to process.In this paper,an improved Euclidean clustering algorithm is proposed for the lack of over clustering or under clustering caused by the inaccurate distance threshold setting in the Euclidean distance clustering method(Edge weighted Euclidean clustering,EWEC)(3)On the basis of segmentation and clustering,this paper proposes a model-based Kalman filter tracking method.Firstly,the tracking object is modeled,then the Kalman filter is used to calculate the prediction and update of the tracking object state.Finally,the matching probability between the registration cluster and the predicted target state between the adjacent frames is calculated through data association to complete the target tracking process.(4)In order to intuitively realize the data of point cloud and the experimental results of algorithm,this paper develops the data receiving and visualization system of lidar based on the rviz system of ROS and OpenGL.The system can receive and release the original lidar data and the data processed by the algorithm,and visualize the data.It greatly facilitates the direct observation and verification of the algorithm effect.