Vulnerable road users in Peruvian cities are badly served by standard vision datasets. The mototaxi — a three-wheeled motorized vehicle that carries a large share of informal transport in Andean cities — appears in none of the big international repositories, and local road markings and traffic lights are faded or non-standard. RICUY (from the Quechua rikuy, “to see; to observe”) is an advanced driver-assistance system built for exactly that gap.

The system runs YOLOv8L-seg instance segmentation, ByteTrack multi-object tracking, monocular distance estimation from a stenopeic camera model, and spoken alerts in Spanish. The experimental core is a transfer-learning and localization study: we start from a YOLOv8L-seg checkpoint pretrained on Ayacucho traffic and fine-tune it on coastal Lima traffic using 253 locally annotated images across 12 classes.

Evaluated on an identical validation split of 464 instances, fine-tuning on local data raises box mAP50 from 0.355 to 0.613 (+72.7%) and mask mAP50 from 0.318 to 0.556 (+74.8%), with simultaneous gains in precision and recall. The most illustrative case is the red traffic light, whose box mAP50 goes from 0.013 to 0.711. The full system was validated on real driving sequences from Lima and Andahuaylas, holding 14–43 frames per second on a 4 GB NVIDIA RTX 3050 after just 41 minutes of fine-tuning — a direct measurement of the localization gap, and evidence that instance-segmentation ADAS is viable for Peruvian urban traffic on modest hardware.

← All research