Learning-Deep-Learning

Robust Traffic Light and Arrow Detection Using Digital Map with Spatial Prior Information for Automated Driving

September 2020

tl;dr: Combines DL and conventional method and HD map prior info for robust traffic light detection.

Overall impression

This paper looks highly practical and engineering focused. It provided many benchmarks to compare against. It is limited to traffic light in Japan, which is perhaps a well regulated market without too many out-of-spec traffic lights. Still, there exist many challenges caused by the nature of the problem, such as distance, lighting condition, etc.

The paper demonstrated flat F1 values up to 150 m (5 pixels).

Key ideas

Technical details

Notes