Development of Obstacle Detection Based on Region Convolutional Neural Network for Autonomous Car

Autonomous car is a transportation technology that has been developed. Its potencies can be run without human operators that decrease the road accident rate. The obstacle detection system becomes one of the significant systems for autonomous cars because it uses for sensing close obstacles. Indonesian people still rarely use the autonomous car because some objects can not be acknowledged by communal autonomous cars like becak. In this research, this obstacle detection system uses a dataset developed for an autonomous car in Indonesia. Faster Region Convolutional Neural Network (F-RCNN) with Residual Network-50 (ResNet-50) and Feature Pyramid Network (FPN) as the backbone system is applied. For training and validation, the self-made dataset comprises 1,451 annotations for the training process and 502 for validating process. The result is good enough which its Average Precision (AP) is 45.67% for 10,000 iterations, 43.07% for 40,000 iterations, and 43.26% for 55,000 iterations. The outputs from the obstacle detection system are an image for visualizing image resulted, the coordinates of objects detected, and their classes for the autonomous car input variables. The result for processing video also shows this system can process the image within ∼ 5 frames per second (fps) with the help of Tesla T4 15,109 MB Graphic Processing Unit and Intel ® Xeon Central Processing [email protected] GHz.