Embedding Object Avoidance to End-To-End Driving Systems by Input Data Manipulation

초록

In this paper, we present a simple, yet efficient data-manipulation method to embed object avoidance feature to end-to-end driving system. Previous works used either additional sensors or extra algorithm to embed object avoidance to end-to-end driving. However, the proposed system tried to keep the simplicity of the end-to-end learning. To this end, we conduct marking on each input image to indicate the location of objects first. This is done by connecting an object-detection network to the front of the steering-estimation network. Thereafter, the proposed steering-estimation network learns the steering angle from the manipulated image sequence. We tested several ways of marking for better understanding of object location. Furthermore, we modified the steering angle estimation network which is based on PilotNet so that the network can estimate the proper steering angle even with the existence of objects. Experimental results show that the proposed network successfully performs object avoidance and steering-estimation accuracy has been improved by 10% compared to PilotNet. Since the proposed system does not require many resources (e.g., millions of data) to perform autonomous driving, we believe it is suitable for systems that perform driving in a specific area: Autonomous valet-parking systems.

키워드