Real-time Obstacle Detection on Embedded System

Human factor causes most of the road accidents; for improving traffic safety, advanced driver assistance system and autonomous driving have become active research. The most significant challenge of these systems is the development of a reliable obstacle detection which can recognize pedestrians and pedestrian-like objects on the road from color images and thermal images. Consequently, the object detection based on neural networks, such as YOLO [1] , is emerging as a solution for identifying obstacles with various appearance. Although the convolutional neural network provides a robust and reliable result, it still involves substantial computational costs and hard to port to an embedded device. This work intends to improve the performance of YOLO and aims at presenting a real-time obstacle detection system on an integrated device.