Research On Road Target Detection System Based On Multi-source Data Fusion

As the most basic and important link in intelligent driving technology,environment perception has become a research hotspot in this technology area.Due to the limitation of hardware conditions,the perception solution based on single sensor has low perception accuracy in some scenarios.With the gradual reduction of radar hardware cost,in order to make smart driving cars better adapt to complex scenes and improve the perception accuracy of smart car perception system for target information in different environments,a fusion detection system based on millimeter wave radar,Li DAR and industrial camera is proposed in this paper.Firstly,the current research status of target detection technology at home and abroad is analyzed and summarized,the hardware characteristics of different sensors are introduced,and the research status of different sensor fusion detection is compared,the overall design of the research system in this paper is carried out,and the corresponding research line is proposed.Then,the parameter models of industrial cameras were determined according to the detection accuracy requirements;the pre-processing of visual data with time stamps and IDs was adopted;an improved YOLOv5 neural network was proposed to change the number of input feature layers in the FPN layer of the traditional YOLOv5 network from three layers to four layers,so that the shallow and deep features can be more fully combined,thus increasing the richness of the extracted information and thus ultimately improve the overall detection accuracy of the target network.Secondly,the parameters of millimeter wave radar and lidar models are determined according to the detection accuracy requirements;the millimeter wave radar data are preprocessed,and the stationary targets are filtered out using the velocity judgment method and cohesive hierarchical clustering algorithm,while the target survivability is judged based on the third-order Kalman filtering method,and finally the surviving targets are output to complete the target detection based on millimeter wave radar;the lidar data are preprocessed with the combination of effective point cloud extraction and filtering processing,and finally the improved DBSCAN clustering detection algorithm is used to complete the target detection based on lidar data.Again,a multi-source fusion perception algorithm framework is proposed.The framework mainly consists of five parts: spatio-temporal synchronization of sensor data,target classification detection based on region of interest(ROI)and intersection ratio of three sensor detection frames,target location detection based on radar data and weighting algorithm,unfused point processing strategy,and target survivability determination based on extended Kalman filter(EKF).In the sensor data spatio-temporal synchronization part,the temporal and spatial coordinates of the three sensor data are unified to carry out the subsequent fusion detection;in the ROI-based fusion target detection part,the region of interest is formed based on the detection information of the two radar data,and then the improved YOLOv5 neural network is used to detect the target in this region to obtain the visual detection data.After the above steps are completed,the target classification detection results are output based on the intersection-and-comparison strategy;in the position-aware part,the data from the two radars are processed using a weighting algorithm t o obtain the position information of the target.After that,the unmatched measurements of the three sensors are processed according to the processing strategy of unmatched points to realize the processing of all measurement points;finally,the EKF is use d to manage the target trajectory,verify its survivability,and output the detection results of the target to be detected by the survivability judgment to complete the whole fusion recognition process.Finally,a multi-source fusion perception experimental platform was built.Firstly,the calibration of the three sensors was carried out,and the real-world experiments were conducted under different road scenarios to collect the relevant data.Secondly,the experimental data were analyzed to verify the feas ibility and accuracy of the multi-source sensing fusion system.