Effectiveness of Focal Loss for Traffic Sign Detection Using Deep Neural Networks

Abstract

Recent developments in autonomous driving have created a great demand for precise and computationally
effective traffic sign-detecting systems. By assisting drivers and assuring their safety, such technology can
lessen traffic accidents and fatalities. However, to make such a system deployable, a few crucial accuracy
and processing performance problems must be resolved. Real-time performance is sometimes regarded as a
must for such an application. RetinaNet, a focal loss-based single-stage object detector, is employed to strike
a compromise between accuracy and processing speed concerning the most cutting-edge object detectors. The
detector is suitable for traffic sign identification since it was developed to overcome the class imbalance problem
that the single-stage detector had. (TSD). The efficiency of the detector was evaluated by combining feature
extractors like ResNet-50 and ResNet-101 on two openly available TSD benchmark datasets. Various metrics
like memory allocation, mean average precision (mAP), running time, amount of floating-point operations, and
model parameters are taken into consideration. Evaluation of the detector on several datasets is required to
examine the variance in the performance, and the RetinaNet model is the fastest and best model in terms of
memory usage, making it the ideal option for the deployment of mobile and affordable embedded devices.