An experimental analysis of artificial intelligence (AI) use for traffic monitoring in urban environments

This article presents an experimental analysis of lightweight convolutional neural network (CNN) object detection models designed for on-device traffic monitoring in urban environments. The study investigates the performance of several mobile-oriented detectors, including EfficientDet, SSD MobileNet V2, and SSD MobileNet V2 FPNLite, with the goal of identifying an optimal balance between detection accuracy, inference speed, memory footprint, and energy efficiency on resource-constrained Android devices. To assess practical applicability, all models were evaluated under realistic operational conditions, including varying object distances, partial occlusions, and reduced illumination typical of urban monitoring scenarios. The comparative analysis shows that the recommended configuration – SSD MobileNet V2 FPNLite (640×640) accelerated with NNAPI – achieves the most favorable trade-off for real-time deployment, reaching approximately 40% mAP on the evaluation dataset while maintaining fast on-device inference and reduced power consumption. Experimental testing further demonstrates that the system achieves up to 94% recognition accuracy at close range and delivers stable performance at medium distances, surpassing several lightweight state-of-the-art detectors in practical real-time tests. Additionally, a modular Android application based on the Model-View-Controller architecture is presented, demonstrating seamless integration of the selected model into an end-to-end mobile processing pipeline. The results confirm that accurate and efficient on-device object detection for traffic monitoring can be achieved without reliance on high-end hardware or cloud-based computation, making the proposed solution well-suited for mobile, embedded, and edge-intelligent urban applications.