Research On Embedded Deep Learning Driverless Safety

With the rapid development in the field of driverless operation,safety and anti-collision technology has become the core subject of current research.The computer vision technology is applied to collect the motion parameters of vehicles or obstacles ahead to achieve the most cost-effective driverless operation scheme.The high accuracy of object detection in vehicle object detection and tracking can be ensured by utilizing convolutional neural network.However,the algorithm has a high demand for hardware resources,and the high-performance hardware is with large volume and in high price.There are many disadvantages in its application of vehicle mobile environment.For the above problems,a small embedded system is used to realize the moving vehicle detection and algorithm tracking of convolutional neural network in this paper.The demand for active safety and anti-collision are met by improving the neural network algorithm,reducing the consumption of hardware resources and optimizing the performance of algorithm.The main contents of this paper are as follows:1.Based on the object detection of moving vehicles,an algorithm suitable for embedded platform is proposed.SSD algorithm is used as the main framework of detection model,and lightweight Mobilenet V2 algorithm is used as the improved framework,so that the detection speed is greatly improved.2.For the part of object tracking,SORT algorithm is proposed to use as the core algorithm,in which the method of data association is used to combine the two parts of technology,and finally realize the real-time vehicle detection and tracking in the embedded end.3.For the embedded system development,RK-3399 chip is used to establish the vehicle detection and tracking system,in which the framework and library needed for vehicle detection and algorithm tracking are transplanted to the embedded end,and the real-time vehicle detection and tracking is finally completed.Experiments prove that the frames per second of the neural network improved by Mobilenet V2 framework is greatly improved for Mobilenet V1 framework,and the accuracy is also greatly improved.By dividing the vehicle detection and tracking into two parts,the first frame of each cycle is for vehicle detection,and the remaining frames of each cycle is for vehicle tracking.In this paper,many experiments have also been done on vehicle detection and tracking cycle to find out the best cycle and further improve the real-time vehicle detection and tracking rate.And the use accuracy is very high,and the frames per second is considerable.From the experimental results,the scheme of vehicle detection and tracking based on RK-3399 embedded driverless operation system is feasible and with a bright prospect.