Traffic Light Detection using Attention-Guided Continuous Conditional Random Fields
Dongwook Yang, Seung‐Woo Seo · 2022
Traffic light detection plays a central role in Advanced Driver Assistance System (ADAS). For an autonomous vehicles to move smoothly and safely on the road, not only it is crucial for an ego vehicle to detect all presented traffic light candidates, but also to not yield any false positives. It is necessary for Convolutional Neural Networks (CNNs) to attend more focus on important features and suppress non-useful details through activations. To carry out such task, we propose a novel network that employs continuous Conditional Random Fields (CRFs) to fuse multi-scale information from different layers of a CNN, guided by attention modules. Extensive experiments are conducted on traffic light detection dataset, which we have acquired with own our autonomous driving platform SNUver. Results indicate that by incorporating attention-guided CRF module inside the network, the network focuses more on regions with traffic lights and thereby improves accuracy and recall rate.