Learning-Deep-Learning

Traffic Light Mapping, Localization, and State Detection for Autonomous Vehicles

December 2020

tl;dr: Mapping and online detection of TFL.

Overall impression

Traffic light state perception is the key.

Even in non-autonomous vehicles, traffic light state detection would be beneficial alerting inattentive drivers to changing light status and making intersections safer. –> TLA

The key to safe operation is the ability to handle common failure cases, such as FP or transient occlusion for camera based TFL perception.

Key ideas

• TFL tracking and triangulation according to Optimal Ray Intersection For Computing 3D Points From N-View Correspondences.
• Temporal filtering with hysterisis: only change light state when several frames of identical color detection have occurred sequentially. This adds a fraction of a second latency but the response time still matches or is better than that of a human.

Technical details

• Camera with fixed camera
• Coordinate transformation: world, vehicle, camera and object
• Traffic light color hue histogram

Notes

• Questions and notes on how to improve/revise the current work