LKTSF-Net: A Small Target Detection Net for Traffic Signal

Despite significant strides in traffic sign detection, effective solutions remain elusive. Traditional models often struggle with low detection accuracy and inefficiency, particularly when dealing with small objects. In response, we present LKSTF-Net, a novel algorithm built on a one-stage object detection framework specifically designed for traffic signal detection. Our approach addresses these challenges by improving detection accuracy and efficiency, making it highly suitable for real-world traffic signal recognition. The core innovations of LKSTF-Net are as follows: First, the LKSTF-PAN feature extraction pyramid utilizes depthwise separable convolution with large kernels, effectively expanding the receptive field while incorporating a dedicated detection head for small objects. This enhances the model’s ability to detect small traffic signals, which is a common challenge in traffic sign detection tasks. Second, the C3TR module, which integrates a classical attention mechanism, captures global information in the feature map and focuses on the most relevant parts of the input, thereby improving overall detection performance. Lastly, we replace traditional Non-Maximum Suppression (NMS) with Soft-NMS, which adjusts the confidence scores of bounding boxes, helping preserve correct detections that might otherwise be discarded by NMS. We validate the effectiveness of LKSTF-Net through comprehensive experiments on the CCTSDB and TT100K datasets. The results demonstrate that LKSTF-Net outperforms other state-of-the-art models in precision (P), recall (R), F1 score (F1), mean average precision (mAP), and model size, confirming its superior performance. These findings highlight the potential of LKSTF-Net for practical traffic signal detection applications, providing a robust and efficient solution for realtime traffic management systems. The codes of this work are available openly at our GitHub repository.