Hybrid LED traffic light detection using high-speed camera

Automatic traffic light detection (TLD) plays an important role for driver-assistance system and autonomous vehicles. State-of-the-art TLD systems showed remarkable results by exploring visual information from static frames. However, traffic lights from different countries, regions, and manufactures are always visually distinct. The existing large intra-class variance makes the pre-trained detectors perform good on one dataset but fail on the others with different origins. One the other hand, LED traffic lights are widely used because of better energy efficiency. Based on the observation LED traffic light flashes in proportion to the input AC power frequency, we propose a hybrid TLD approach which combines the temporally frequency analysis and visual information using high-speed camera. Exploiting temporal information is shown to be very effective in the experiments. It is considered to be more robust than visual information-only methods.