AI-Powered Vehicle Counting and Classification in Smart Cities

Abstract

The rapid growth of urban traffic has outpaced the capabilities of traditional loop sensors and manual surveys, creating an urgent need for scalable, low-latency, and cost-effective traffic intelligence. This manuscript presents an end-to-end, AI-powered system for vehicle counting and classification designed for smart-city deployments. The pipeline integrates single-shot object detection with multi-object tracking to produce de-duplicated counts and fine-grained classes across heterogeneous camera views. The proposed method emphasizes practical constraints: camera placement variability, day–night domain shifts, adverse weather, heavy occlusions, and compute limits on edge devices. We fuse a one-stage detector (for bounding-box localization and coarse class labels) with an appearance-embedding tracker to maintain identities through occlusions and support virtual line-crossing logic that yields robust counts. An optional attribute head refines classes (e.g., car, bus, truck, two-wheeler, auto-rickshaw) using shape priors and aspect ratios.

We also introduce normalization techniques (perspective-aware regions of interest, homography-based scale cues, and temporal smoothing) that stabilize predictions under viewpoint changes. Simulation-based evaluations (CARLA + SUMO) emulate dense intersections with configurable lighting and weather, producing 100k labeled frames across five junction archetypes. The system attains high detection accuracy (mAP@0.5 = 0.81), strong tracking (IDF1 = 0.78), and reliable counts (overall MAE = 2.3 vehicles/minute lane-crossing) at 25–30 FPS on an NVIDIA Jetson-class edge device via INT8 quantization. A statistical analysis demonstrates consistent performance across classes and time-of-day, with night-time recall improved by temporal voting. The results suggest the approach is deployable at city scale, enabling real-time traffic planning, adaptive signal control, and safety analytics with modest infrastructure upgrades.