Development Of Dynamic Target Detection And Ranging Device Based On Jetson

With the rapid development of science and technology in recent years,people’s requirements for product’s intelligence and autonomy are getting higher and higher.In the field of autonomous driving and assisted driving,the environmental perception and autonomous decision-making capabilities of actors have become the focus of attention.Target detection and distance measurement technology is the key technology,which profoundly affects not only the ability of the products to tell the different types of targets in the surroundings,but also the distance perception capability.Target detection and ranging technologies based on single sensor and traditional target detection algorithms are increasingly unable to meet the subject’s requirements for perception of complex environments.The target detection algorithm based on deep learning and the environment perception based on sensor fusion can improve people’s ability to perceive the environment in complex scenes.Aiming at the environment perception detection function in the field of autonomous driving,this thesis adopts the environment perception technology of sensor fusion and the target detection algorithm based on deep learning to develop a set of target detection and ranging device based on the embedded platform,which has high practical value.This thesis uses sensor fusion technology combined with the help of deep learning algorithms to realize the target detection and tracking and ranging function.The equipment is divided into four parts: sensor data acquisition,spatial position calibration,data frame synchronization,and target detection algorithm optimization.Through investigating and comparing of embedded heterogeneous platforms and external environment sensing sensors,this thesis has finished the selection of hardware platform and sensors and the formulation of software implementation schemes were carried out,and sensor data collection and preprocessing were completed.In order to solve the problem that the YOLO-V4 target detection algorithm runs poorly on the Jetson platform,a research on algorithm optimizating acceleration was carried out.The backbone network of the Mobile Net-V3 algorithm was introduced to replace the network backbone of YOLO-V4,and the VOC data set was used to complete the training of the model under Pytorch.The optimized model traded for higher running speed and smaller model size with minimal loss of detection accuracy.In order to obtain the three-dimensional space information of the target,a three-dimensional space fusion model combining millimeter wave radar and camera data is constructed.In the case of millimeter-wave radar filtering and image distortion correcting,the target spatial position fusion and time synchronization of radar and image data frames are realized.After experimental verification,in outdoor scene,the embedded target detection and distance measurement system can accurately tell the different types of objects,realize the target ranging function,and provide the products with accurate surrounding environment perception information,which meets the technical requirements.