Research On Roadside Point Cloud Target Detection And Application Based On Lightweight Networ

In the intelligent transportation scenario,roadside sensing devices collect and process various environmental data,interact with intelligent cars,and promote vehicle-infrastructure cooperation.Point cloud object detection can process the 3-D spatial information of the road environment and is a crucial technology in roadside perception.At present,although the detection accuracy of most point cloud object detection methods has continuously improved,the computation and parameters of the network have also increased.It is difficult to deploy to roadside equipment,and the detection efficiency is low,so it can not meet the real-time requirements.The application is complex.Therefore,this thesis takes the roadside application of point cloud object detection as the starting point and carries out roadside point cloud object detection and application research based on a lightweight network based on the characteristics of roadside scenes and actual application requirements.The main work and innovations of this thesis include the following aspects:(1)This thesis constructs a roadside point cloud dataset.Due to the lack of public point cloud datasets for the roadside scenes,this thesis designs a roadside point cloud collection scheme based on the characteristics of roadside scenes,builds a collection device,and finally completes the construction of the roadside point cloud dataset.(2)A one-stage point cloud object detection network(Point Pillars-MFA)based on multilayer multi-scale feature fusion and 3-D attention is proposed.In the model,this thesis introduces a multi-layer feature pyramid structure and a 3-D attention mechanism,and presents a multi-scale feature extraction module that combines 3-D attention for feature enhancement.Experimental results show that Point Pillars-MFA has achieved higher detection accuracy in multiple categories.(3)A lightweight roadside point cloud object detection network(Light-Point Pillars-MFA)based on context enhancement and Ghost module is proposed.The network is lightweight and improved based on Point Pillars-MFA: First,this thesis introduces a lightweight context enhancement structure to simplify the multi-scale feature extraction module to reduce the number of parameters and calculations;then,the Ghost module is presented,which replaces the more efficient Ghost Convolution with standard convolution.The experimental results show that Light-Point Pillars-MFA has fewer parameters and calculations,improves the detection accuracy and speed,and has a high roadside application value.(4)This thesis designs and implements a roadside point cloud object detection application system based on an edge computing unit.For the actual roadside application scenarios,the system deploys the Light-Point Pillars-MFA network to the edge computing unit.It performs model acceleration based on Tensor RT to improve the detection speed further.Experiments in the roadside show that Light-Point Pillars-MFA achieves higher detection efficiency of roadside point cloud objects after Tensor RT acceleration.