The deep learning techniques have been shown to make a traffic objects classification system for V2V communications to ensure traffic safety and traffic flow prediction. The robust classifier on the base of MLP and PointNet are explored to recognize the traffic objects from lidar point clouds. The features of a lidar sensor, the lidar point cloud coordinate system and its complex properties for creation a smart traffic object detection and recognition model are described. The best configuration of PointNet architecture with hyperparameters are shown, which is more efficient and robust with respect to input perturbation and corruption of lidar point clouds.
The deep learning techniques on the base of YOLOv5 architecture have been shown to make a traffic objects detection system for V2V communications to ensure traffic safety and traffic flows prediction. According to the detection data the detector allows to estimate the traffic object speed and the coordinates of movement. To improve the accuracy of calculations, the distortion correction algorithm (in geometric optics), the Chang method and the Kalman filter are used. To set the detector to work in online operation mode, an algorithm has been developed that receives video stream from open city surveillance cameras with the processing of a m3u8 playlist. The developed detector can be used on "connected" vehicles to assess the situation on the roads and ensure safety traffic.