Research Of Urban Road Vehicle Trajectory Tracking And Automatic Generation Algorithm Based On Deep Learning

The traditional vehicle-oriented testing methods have been unable to meet the requirements of autonomous vehicle testing,which is a strong coupling system of vehicle-road-environment.Scenario-based virtual testing methods have become an important testing method for autonomous vehicles due to its advantages of flexible scenario construction,high test efficiency and good test repeatability.Therefore,the construction of real and effective testing scenarios has become the key to carry out virtual testing of autonomous driving.The urban road traffic scenario is complex and changeable,which is characterized by high dimension,randomness and time continuance,including multi-dimensional heterogeneous scenario elements such as pedestrians with various behaviors and postures,vehicles with different movement tracks,road structures with different forms and traffic signs with different spatial positions.According to the characteristics of urban road traffic scenario,compared with the traditional scenario construction method based on expert experience,the scenario construction method based on real traffic data can ensure the authenticity and effectiveness of the testing scenario.In the construction of virtual testing scenario,roadside video surveillance data can provide rich urban traffic scenario data as the data source for the construction of autonomous driving test scenario.In this paper,vehicle trajectory is taken as the main research object in the construction of test scenario.A series of key problems such as vehicle trajectory detection,tracking and generation are solved,and a diversified,high quality and realistic vehicle trajectory generation model for complex urban roads is built.The main research contents of this paper include the following four aspects:(1)For the video data of traffic scenario,the vehicle detection model and vehicle tracking model were built respectively in this paper.Based on the characteristics of urban traffic image data,this paper builds the Faster R-CNN object detection network to extract the position information of moving vehicles in the traffic scenario video.Then,based on the results of vehicle detection and the characteristic information of vehicles in the video image,the Deep SORT algorithm based on the characteristics of the dataset in this paper is built,and to extract the trajectory of moving vehicles in the traffic scenario,and then 25 urban road vehicle trajectory datasets can be obtained.(2)Two constructed methods of vehicle trajectory datasets were proposed,namely,vehicle trajectory image data and vehicle trajectory sequence data.The main purpose of the construction method of vehicle trajectory image data is to express indirectly the position information and velocity information of the vehicle tracking using the color features of the image,after the vehicle trajectory is preprocessed and color converted,which can be used for the GAN training,and that is made of two-dimensional convolution layers.And,the main purpose of the construction method of traffic trajectory sequence data is to directly represent the position and velocity information of the vehicle trajectory with the one-dimensional sequence data of fixed length,after the vehicle trajectory has been preprocessed and sampled at key points,which can be used for the GAN training of one-dimensional convolution layers.(3)A T-WCGAN model based on vehicle trajectory image data was bulit.For the vehicle trajectory image datasets,the trajectory generation model is considered with combining with the principle of CGAN,DCGAN and WGAN.An RGB image generator and an RGB image discriminator composed of a two-dimensional convolutional layer were constructed,and then the T-WCGAN model based on the traffic trajectory image data was formed.(4)A T-WCGAN model based on vehicle trajectory sequence data was bulit.For the vehicle sequence image datasets,a trajectory sequence generator and trajectory sequence discriminator composed of one-dimensional convolutional layer were construced in this paper,and then the T-WCGAN model based on the traffic trajectory sequence data was formed.Finally,for those two T-WCGAN models based on trajectory image data and trajectory sequence data,the tests of generated vehicle trajectory visualization and comparative analysis were conducted in this paper.Both the image-based T-WCGAN model and sequence-based the T-WCGAN model have achieved the expected results and performance performance in the experiment,both of those two T-WCGAN models can input some dynamic vehicle trajectory(location information and vehicle information of the trajectory).For the perspective of the details of generated trajectory,the sequence-based T-WCGAN model is better than the imagebased T-WCGAN model.