Navigating Self-Driving Vehicles Using Convolutional Neural Network

Engineering notes

With the combination of three cameras, the accuracy of this navigation task is improved significantly.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

In this paper, a method for navigation of self-driving vehicles is proposed. Although the research for this problem has been performed for several years, we noticed that the elevated accuracy results have not been achieved yet. Therefore, the method using a convolutional neural network (CNN) for training and simulation of unmanned vehicle model on the UDACITY platform has been made. Details, we used three cameras mounted in front of a vehicle to follow three directions were left, right and center position to collect data. The data are the images that captured from three cameras. The number of samples image is 15504. In this research, the label with two parameters are the steering angle and speed from each image would also be created. After collecting the data, these parameters will be achieved by training CNN used to navigate the vehicle. With the combination of three cameras, the accuracy of this navigation task is improved significantly. When vehicle deviates to the left, we will compute the error of the steering angle value between the middle and left position. Afterward, the steering angle value will be adjusted to control the vehicle could run in the center of the lane. Similarly, in the case when vehicles deviate to the right. Based on the simulation platform of UDACITY, we simulated and obtained the result with accuracy was 98, 23%.

Links and sources

• Official / arXiv page

Comments