Vulnerable road users in informal urban environments confront a distinct set of hazards that standard computer vision datasets are ill-equipped to represent: artisanal speed bumps constructed without regulatory compliance, deteriorated road markings, and the mototaxi—a three-wheeled motorized vehicle that constitutes the primary informal transport mode in intermediate Andean cities yet is absent from all major international repositories. This paper presents QHAWAY—from Quechua qhaway, a transitive verb meaning “to look; to observe”—an Advanced Driver Assistance System (ADAS) predicated on instance segmentation, monocular distance estimation via the pinhole camera model, and Time-to-Collision (TTC) computation, developed for the road environment of Ayacucho, Peru (2761 m a.s.l.), a city recognised by UNESCO as a Creative City of Crafts and Folk Art since 2019. A hybrid dataset comprising 25,602 images with 127,525 annotated instances across 12 classes was assembled by combining an original local collection of 4598 images (10,701 instances) captured through four complementary acquisition methods across the five urban districts of the Huamanga province with three established international datasets (BDD100K, BSTLD, RLMD; 21,004 images, 116,824 instances). A three-phase progressive training strategy with monotonically increasing resolution (640, 800, and 1024 pixels) was evaluated as an ablation study. A multi-architecture comparison spanning YOLOv8L-seg and the YOLO26 family (nano, small, large) identified YOLO26L-seg as the best-performing model, attaining mAP50 Box of 0.829 and mAP50 Mask of 0.788 at epoch 179. The integration of ByteTrack multi-object tracking with the pinhole equation $D=(H_{\mathrm{real}}\times f)/h_{\mathrm{px}}$ delineates operational risk zones aligned with the NHTSA forward collision warning standard (danger: <3 m; caution: 3–7 m; TTC threshold ≤ 2.4 s). The system sustains processing rates of 19.2–25.4 FPS on an NVIDIA RTX 5080 GPU. A systematic field survey established that 96% of the audited speed bumps fail to comply with MTC Directive No. 01-2011-MTC/14, constituting the first quantitative record of informal road infrastructure non-compliance in the Andean region. Validation was conducted under naturalistic driving conditions without staged scenarios. Grad-CAM explainability analysis, encompassing three complementary visualisation algorithms (Grad-CAM, Grad-CAM++, and EigenCAM), confirmed that model attention concentrates consistently on safety-critical objects.

This paper presents a distributed edge computing system for real-time urban traffic monitoring that integrates vehicle speed estimation and automatic license plate recognition (ALPR) on NVIDIA Jetson Orin NX devices. The system addresses the need for low-cost, automated traffic enforcement in developing urban environments where certified radar-based speed measurement infrastructure is not available. The proposed architecture consists of two independent processing nodes. The first node performs speed estimation using an IMOU aerial camera, YOLO26n-based vehicle detection, ByteTrack multi-object tracking, perspective homography transformation, and Kalman filtering for instantaneous speed computation. The second node performs license plate recognition employing a frontal camera, YOLO26n detection, and a Compact Convolutional Transformer (CCT) OCR model for real-time plate reading. Both subsystems are deployed on Jetson Orin NX modules with TensorRT FP16-optimized models, sustaining 20.0–22.3 FPS for speed estimation and 29–36 FPS for license plate recognition. The system incorporates the Peruvian traffic regulation (DS 025-2021-MTC) for automated violation classification, including speeding offenses (code M20) with the 5 km/h camera tolerance threshold introduced by DS 011-2026-MTC. Field validation was conducted at Jr. Carlos F. Vivanco, Ayacucho, Peru (2719–2725 m a.s.l.). Speed estimation was validated in a 40-pass nocturnal campaign against 1 Hz GNSS Doppler velocity spanning 16.9–48.8 km/h, a range that covers both the 30 km/h statutory limit and the 35 km/h enforcement tolerance. The configuration deployed during the campaign, which smooths the metric foot-point with a constant-velocity Kalman filter, attained a mean absolute error of 3.19 km/h at the enforcement line. The residual was identified as the steady-state lag of that motion model under vehicle acceleration, with which it correlates at r = −0.727; replacing it with a constant-acceleration model reduces the error to 1.32 km/h under leave-one-pass-out cross-validation and removes the dependence on acceleration entirely (r = +0.003). The improvement was verified to transfer to the Jetson deployment node. Two earlier campaigns are retained for completeness because they identified deployment failure modes—an RTSP stream resolution renegotiation and a speed-display defect affecting vehicles stopped beyond the measurement zone—that are reported as engineering contributions in their own right. A comprehensive model development study, comprising six training experiments (M1–M6) across three architectures (YOLOv8m-seg, YOLO11m-seg, YOLO26-seg/det), documents the methodological transition from instance segmentation to bounding-box detection driven by edge computing latency constraints. The results demonstrate the feasibility of deploying a traffic enforcement prototype using commercial edge devices and IP camera infrastructure, offering a scalable and cost-effective solution for smart city applications in developing regions.