Urban traffic monitoring based on deep learning on an embedded GPU

This paper presents a deep learning-based system for urban traffic monitoring, focusing on the detection and tracking of motorcycles using embedded hardware, due to the high accident rates of this type of vehicle. Different convolutional neural network (CNN) models were evaluated, including MobileNet-v1-SSD, YOLOv5, and Faster R-CNN, implemented on an NVIDIA Graphics Processing Units (GPUs) board as the Jetson Xavier NX®. The MobileNet-v1-SSD model stands out for its balance between precision (90 %), recall (66 %), and latency (∼10 ms), making it ideal for real-time applications. Additionally, a tracking algorithm based on optical flow using the Lucas-Kanade method was developed, complemented with logic for creating and deleting identities (IDs), enabling object tracking in dynamic scenarios with partial occlusions. The system includes a methodology for calculating key traffic variables such as speed and direction by correlating pixels with real-world distances through camera calibration. This approach demonstrates the feasibility of developing complex image-processing applications based on resource-constrained platforms by leveraging the features of efficient embedded systems such as General Purpose GPUs.