Comparative Analysis of YOLOv4 and EfficientDet based models for Traffic Sign Detection in Autonomous Vehicles

Visual perception systems enable autonomous vehicles to make right driving decisions in dynamic driving environments. Traffic sign recognition is an important perception application, where accuracy and real-time execution are key requirements. Deep learning models like YOLO and RetinaNet have shown reasonably good accuracy and speed as detectors in traffic sign recognizers. With compound scaling, EfficientDet models of varying complexity and performance can be created for use in systems with different computational power. This paper presents a comparative analysis on the usability of YOLOv4 and EfficientDet models in the detection stage of traffic sign recognizers. Both models are trained and evaluated using GTSDB dataset. Comparative performance analysis of YOLOv4, EfficientDet model family and prior models like Faster-RCNN, YOLOv3 and RetinaNet, for traffic sign detection, is reported in this paper. YOLOv4 with a mAP of 97.7@9.18fps is found to be the most suitable one as traffic sign detectors in autonomous vehicles.