Transformer Fusion and Residual Learning Group Classifier Loss for Long-Tailed Traffic Sign Detection

Automatic traffic sign detection and recognition (ATDR) plays a crucial role as a submodule of the intelligent transportation system (ITS) and has gained significant attention in recent years. However, ATDR faces challenges due to two main factors: the small size of traffic signs in images and the long-tailed distribution of sign categories in the real world (i.e.,only a few categories have sufficient samples, and most categories have insufficient samples). These challenges result in a decline in the performance of existing detection frameworks. In this article, we propose a novel framework for traffic sign detection, specifically designed to tackle the challenges posed by these issues. Our approach includes an efficient adaptive and transformed spatial feature fusion module (ATS2F) for detecting small traffic signs. This module adaptively learns feature map associations to extract and fuse meaningful features through cross-space and cross-scale interactions at different levels. In addition, to tackle the long-tailed traffic sign problem, we have designed a residual learning group classifier loss (RLGCL), by dividing the traffic signs into two levels, reserving a specific capacity for the long-tailed category from the perspective of the parameter space, the residual fusion mechanism is introduced to enhance and optimize the tail category image. Our experimental results, obtained from rigorous evaluations on the Tsinghua-Tencent 100K (TT100K) and GTSDB datasets, demonstrate the effectiveness of our approach. The improvements achieved in small traffic sign detection and long-tailed detection underscore the significance of our framework in advancing ATDR research.