Development of Single-board Computer-based Self-Driving Car Model using CNN-Controlled RC Car

Engineering notes

Key topics: self-driving car, self-driving, prediction, control. See the paper for implementation details and experimental results.

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

One of the critical stages in developing an autonomous car is to test the AI controller model using an autonomous car prototype. However, the actual size prototype cars are quite expensive and may be dangerous if the fault in controlling action has occurred. Therefore, this research aims to prototype a small-scale autonomous car using an RC car with a single-board computer. AI is built by using the convolutional neural network (CNN) method. The camera images were used as sensor input (training step), and the steering wheel angle and the car's speed as output. The first stage of this method is to collect training data carried out by recording images, steering angles, and car speed when the car is running on its track. The amount of data taken is 3000, 6000, 12000, and 24000 data. Another step is to perform deep learning training to the model, which has variations in convolution layers as 2, 3, 4, and 5. The next stage is testing the prediction of steering angle and car speed from image data using the trained model with lighting and the color of the obstacles varied. The test found that the model with 24000 data and three convolution layers produced the slightest absolute prediction error at 0.18257.

Links and sources

• Official / arXiv page

Comments