Learning deep feature fusion for traffic light detection

Traffic light detection in real-world conditions is challenging because of the positioning of lights, variety in shapes and scales, and similarity with other objects. The paper presents a deep learning-based traffic light detection system by learning the fusion of handcrafted features. The handcrafted features for object detection focus on specific attributes such as shape, color, or texture. The objective of this work is to incorporate handcrafted features into the network learning process such that the resulting detector parameters are robust to input variations, sensor noise, and atmospheric noise. The proposed detection framework is based on the latest You only look once (YOLO) architecture, trained with the fusion of different information channels in the Integral Channel Features (ICF). The approach demonstrates a qualitative approach for identifying the optimal layer for additional feature injection in the network, and the selection of ICF channels to be applied for fusion. The validation of the proposed detector on the Bosch small traffic light dataset achieved the best mAP score of 55.7 % on the testing set. Further, a qualitative comparison of the proposed detector's performance with that of other recent methods is presented, along with an analysis using auxiliary experiments.