Enhancing traffic management with embedded machine learning for vehicle detection

Abstract

In recent years, vehicle detection has become vital for applications ranging from autonomous driving to traffic control, surveillance, and monitoring. The demand for efficient real-time detection systems has surged, prompting the integration of machine learning algorithms into embedded platforms as a promising approach. This paper focuses on developing a robust and efficient system deployable on the NVIDIA Jetson Nano 2GB Developer Kit. The system harnesses machine learning algorithms tailored for resource-constrained embedded systems to achieve high detection accuracy in realtime. The process encompasses data preparation, preprocessing, feature extraction, and classification. Deep learning models used are You Only Look Once (YOLO) algorithm, YOLOv5n, and YOLOv7-tiny, trained on labeled datasets to classify regions of interest based on unique vehicle attributes. For inference on the Jetson Nano, both are chosen for their real-time capabilities and high object detection accuracy, are employed. To leverage the Jetson Nano's GPU power, NVIDIA's Compute Unified Device Architecture (CUDA) toolkit is installed, enabling parallel computing and deep learning model optimization. Results indicate that YOLOv7-tiny achieves the better precision-confidence at 92.7% and a recall-confidence of 94% compared to YOLOv5n. This study also uses various other evaluation metrics such as accuracy, precision, recall, confusion matrix, and F1 score to measure system performance and examine computational efficiency, to help in the selection of appropriate models for embedded systems.