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
- Combine Yolo, SURF keypoint and blob object detection.
- SURF: detecting blobs by det(H)
- HD Map should contain info about traffic light
- 2D TL position (lat, long, heading orientation), 3DoF. –> Only traffic light within a certain range and orientation to ego vehicle are projected back to image.
- format (horizontal, vertical, 1x3, 3x2, etc)
- No height information, only approximate height info during online inference according to country (~5.0 meter in Japan) or set a wide recognition area in image when height is unknown.
- Highlighted image generation in HSV space
- Normalize the brightness Value to emphasize the lighting
- Update Saturation to eliminate background noise
- Weighting wrt Hue color
- Arrow recognition
- One detector is trained to recognize right arrow only. For left/straight arrows the image is rotated.
- The paper also incorporates prior information (right arrow under red light) into arrow recognition. –> but the paper did not see how.
Technical details
- Deceleration without discomfort to passengers is approximately 0.1 G. Recognizing TFLs in the ranges exceeding 100 m is required to make a natural intersection approach in automated driving.
Notes
- Questions and notes on how to improve/revise the current work