Real-time dense traffic detection using lightweight backbone and improved path aggregation feature pyramid network

An informed traffic situation at a grade crossing is essential for traffic management. Current detection systems are expensive in computation and unsatisfactory in dense traffic instance detection. This work proposes a lightweight dense traffic detection network (DTDNet-Lite) for improved detection performance to address the issues above and benefit the development of the portable traffic monitoring system, especially at the railroad -highway grade crossing areas. An improved path aggregation feature pyramid network (iPAFPN) is developed for multiple-scale feature fusion. A lightweight backbone, ResNet18, is employed to extract features efficiently and accurately. Comprehensive experiments have been conducted on the VOC 2007 dataset and our customized grade crossing dataset. Results indicate the superiority of DTDNet-Lite, paving the way for the deployment of efficient embedded artificial intelligence (AI) computing devices for better traffic monitoring at grade crossings.

Automated summary

In order to improve traffic management and the performance of the traffic monitoring system at railroad-highway grade crossings, this study proposes a lightweight dense traffic detection network (DTDNet-Lite). The network utilizes an improved path aggregation feature pyramid network (iPAFPN) for multi-scale feature fusion, and employs a lightweight backbone, ResNet18, for efficient and accurate feature extraction. Comprehensive experiments on the VOC 2007 dataset and a customized grade crossing dataset show the superiority of DTDNet-Lite, paving the way for the deployment of efficient embedded AI computing devices for better traffic monitoring at grade crossings.