| With upgrading of electrification and intelligence in automotive industry,unmanned driving,as the main research hotspot in this field,can provide protection for traffic safety and efficiency.Therefore,it has become the main trend of the reform.Unmanned driving involves many technologies such as sensor,artificial intelligence(AI)and computers,especially relying on sensors.Due to the directivity and strong anti-jamming capabilities of laser,lidar can obtain the target’s geometry and distance.It has obvious advantages in response speed and detection accuracy,so it has been widely used in unmanned driving.However,lidar still has maj or problems in cost,resolution and range.With the development of AI,methods based on deep learning(DL)in improving the 3D point and target detection have also emerged.However,these methods still need to improve both in efficiency and accuracy.Considering the problems above,this paper focuses on study of a high-resolution lidar imaging radar,3D point super-resolution reconstruction based on generated adversarial networks,and research on accurate detection of 3D targets based on complex YOLO framework,details are as follows:(1)Considering the shortcomings of lidar in cost,limited range,imaging resolution and detection accuracy,this paper analyzes and studies a high-resolution laser imaging radar.Firstly,it analyzes lidar’s principle and mathematical model.Secondly,it describes the overall scheme,including the main components,key technical indicators and principles of sub-modules.Finally,it illustrate the experiments and test results of laser imaging radar under different actual conditions.The center wavelength is 905nm,the farthest imaging distance>250m and field of view is 45°×45°.The range resolution reaches sub-centimeter level and imaging resolution is 512×512.(2)In order to solve the problems of sparse,uneven distribution and noise with original 3D points,this paper carries out the study of super-resolution reconstruction of sparse points based on the generation of confrontation network.The whole structure of the PSR-Net in this paper is divided into two parts:a generator and a discriminator.The generator performs points feature extraction,feature expansion and upsampling generation on the input points and outputs the reconstruction results.The generator evaluates and predicts the results of generator.In this paper,the connection between feature extraction units in the original model is changed to a more efficient skip connection,which can achieve points feature sharing to ensure the reconstruction accuracy and improve the processing speed at the same time.Through the comparison of multiple sets of experiments,the proposed method has better performance both in reconstruction accuracy and processing speed.(3)Considering the problems of low accuracy and poor performance of small targets in 3D detection in unmanned driving,this paper carries out research based on complex YOLO.With the method Complex-YOLO,this paper implements a 3D target detection model ERPD-Net which is more suitable for smaller target such as pedestrians and cyclists.Based on the multi-view method,the model maps the 3D points into fixed grid units to RGB images,then inputs those to CNN to extract features.With the feature maps,E-RPN network(the traditional method RPN with addition of complex angle regression)can output the target’s position,size,detection probability,classification and so on.Considering the importance of detecting smaller objects such as pedestrians and cyclists in unmanned driving,this paper adds a scale factor to the term related to the size of bounding box in loss function.Through comparative analysis of multiple experiments,the proposed method is proved more adaptable to smaller target in 3D point target detection. |
PDF Full Text Request