Abstract:

An improved Faster-RCNN algorithm for detecting small traffic signs was proposed， which addresses the issues of poor recognition performance of distant small targets and high computation cost in real-world traffic scenes affected by weather and lighting conditions. Based on the basic architecture of Faster-RCNN， the algorithm reconstructs the backbone network and improves the region proposal network to make the network framework lightweight. A multi-scale feature fusion network is designed by integrating the scSE attention and GSConv modules， and the Anchors box size was updated to improve the localization and recognition of traffic sign targets. The ROI Align pooling operation with bilinear interpolation for each target subregion was used to preserve the detailed features of the target region and improve the ability to capture details of distant targets. The balanced L₁ loss function was adopted to address the problem of imbalance between samples with large gradient difficulty and those with small gradient easiness， thus improving the training effect. Experiments were conducted on the expanded TT100K dataset. Results show that compared with traditional Faster-RCNN， the model weight is reduced by 200 MB， and detection accuracy is improved by 21.3%. The algorithm achieves a detection accuracy of 85% in low-intensity environments such as cloudy days， which helps improve the traffic sign detection performance in extreme environments.

Key words: environmental awareness, traffic sign detection, TT100K, Faster-RCNN, multi-scale feature fusion