Single-Camera AI-Based Traffic Analytics for Heterogeneous Intersections: A Deep Learning Framework for Microscopic Data Extraction from Smartphone Video

Abstract

In countries like Nepal, where traffic is heterogeneous, mixed, and non-lane-disciplined, the data collection methods used in transportation planning remain outdated and incapable of producing the microscopic parameters needed to understand local driving behaviors. This paper demonstrates a computer vision framework that converts footage from a single elevated smartphone camera (4K, 24 fps) into microsimulation-ready traffic data at the Koteshwor intersection, a busy three-legged junction in Kathmandu. The framework deploys YOLOv8x for vehicle detection, ByteTrack for multi-object tracking supported by a three-tier identity recovery mechanism, and per-approach planar homography, along with a post-processing pipeline for axial-length reclassification, speed correction, Wiedemann 74 safety-distance extraction, and complementary acceleration/deceleration metrics. From two independent peak-hour datasets yielding 31,882 total vehicle trajectories, 29 stratified validation clips covering 1,805 manually counted vehicles resulted in a 100% GEH pass rate and a class-wise Mean Absolute Percentage Error (MAPE) of 2.48%. The framework outputs volumetric counts across seven vehicle classes, Origin-Destination (OD) matrices, speed distributions, and trajectory-level microscopic parameters, demonstrating that a single smartphone setup can produce richer and more reliable traffic data than traditional manual surveys.