Research On Pedestrian Detection In LiDAR Data

With the rapid development of LiDAR technology and the research on point cloud data processing,the LiDAR system has been widely used in various fileds and plays an increasingly important role in people’s daily lives.Especially with the vigorous development of autonomous robots,autonomous driving,unmanned aerial vehicle(UAV)and other fields,the demand for intelligent and automatic sensing of the environment and is increasing day by day.Because it has the ability to actively acquire high-precision 3D point clouds,LiDAR has become one of the most important sensors for unmanned systems.Pedestrian is the most common and the most important dynamic object in indoor and outdoor environments.Detecting people is crucial for many areas such as autonomous driving/smart driving,autonomous robots,monitoring systems,and intelligent transportation systems.At present,although the pedestrian detection method based on vision sensor has been very ideal in many scenes,it still has many limitations,such as the inability to accurately perceive the location of the object,the vulnerability to the influence of weather and light.Therefore,research on pedestrian detection in LiDAR data is not only a beneficial supplement to the vision-based pedestrian detection under extreme conditions,but also the basis of subsequent research on pedestrian motion estimation,tracking,etc.,which is of great importance to combine multiple sensors to solve environmental perception problems.At present,existing methods have been able to accurately detect pedestrians at close range,with dense point cloud and no occlusions.However,in complex scenes,current methods often show some problems of poor segmentation results and poor description ability of features,and suffer from low detection precisions when facing the interactions of various factors,such as the noise of the point cloud,the uneven density of the point cloud,the complexity of human body shape(the shape,posture,dress,and the dynamic change)and complex interactions between pedestrians and environments and objects.In order to improve these problems and achieve higher precision pedestrian detection in lidar data,this paper starts from the following aspects:(1)Aiming at the problems of over segmentation and under segmentation in the existing methods,we proposed optimal segmentation method based on grid elevation graph for pedestrians detection.The proposed method is vertified and compared with other existing methods in the data of multiple scenarios.The results show that the proposed method can effectively improve the accuracy of pedestrians point cloud clusters.(2)We summarize the features of the existing methods for describing two-dimensional/three-dimensional pedestrian point cloud clusters,and propose a feature mining method based on fusion feature space with these features.Finally,the more robust 2D/3D features(parameters)for pedestrian(leg)detection can be obtained by statistically analyzed based on the experiment results.(3)Aiming at the problem of insufficient description ability of existing features for complex two-dimensional point cloud cluster,we defined and proposed a variety of new feature types and feature parameters.Finally,the experimental results of multiple data sets under multiple segmentation methods show that the pedestrian leg detector generated by the Real AdaBoost algorithm based on the multi-type feature space in this paper can achieve the high-precision detection of human leg.The validity and robustness of the features proposed in this paper have been verified.(4)We focus on solving the problem of insufficient description ability of existing features for complex 3D human clusters.On the basis of the experiences of feature mining and extraction,we defined and proposed new feature types with a series of multi-dimensional feature parameters from zero dimension to three dimension.Finally,the results of several experimental data sets show that the pedestrian detection model generated by SVM with the proposed multidimensional and multitype features can achieve a higher accuracy of pedestrian detection.The effectiveness and superiority of the proposed features in this paper have been verified.In general,this paper mainly optimizes pedestrian detection methods from two key problems of point cloud segmentation and feature description and extraction.According to experimental results,the proposed method can effectively improve pedestrian detection accuracy in LiDAR data,especially in complex scenes.