RTL-Net: real-time lightweight Urban traffic object detection algorithm

Abstract

Object detection algorithm in urban traffic using remote sensing images often suffers from high complexity, low real-time performance, and low accuracy. To address these challenges, we propose RTL-Net, an urban traffic object detection network structure based on You Only Look Once (YOLO) v8s. To enhance real-time performance beyond the benchmark, we implemented lightweight designs for the loss function, backbone, neck, and head components. Firstly, a Powerable-IoU (PIoU) loss function was introduced to make the algorithm more suitable for different scales of targets and reduce false detection. Secondly, the Lightweight Shared Convolutional Detection (LSCD) head was replaced to ensure the detection performance and significantly improve the lightweight performance of the algorithm. Additionally, this paper introduces the Dilatation-wise Residual (DWR) module to facilitate the algorithm’s extraction of detailed features. In addition, we optimize the Bidirectional Feature Pyramid Network (Bi-FPN), enabling the fusion of multiple features to improve overall feature integration and performance. The VisDrone2021 dataset was utilized for experimental training. Experimental results demonstrate that the proposed algorithm achieves a significant 43.9% reduction in parameters and an 18.9% decrease in computational complexity. Moreover, the detection accuracy has improved by 2.3%, while maintaining a real-time detection speed of 263.2 frames per second. For edge computing object detection, our method outperforms YOLOv8s and leading remote sensing algorithms in both speed and accuracy, achieving state-of-the-art performance among single-stage detectors with comparable parameters.