India operates one of the largest road networks in the world, and every kilometre of it depends on a precise inventory of road safety features - crash barriers, signs, lane markings, delineators, lighting, and pavement condition - to keep drivers safe and operators compliant with IRC and MoRTH standards. Auditing those features today is overwhelmingly manual: survey teams drive routes with clipboards and cameras, log conditions by hand, and produce static reports days or weeks after the inspection.
Manual highway audits have hit a scalability ceiling. With over 150,000 road fatalities annually and rising regulatory expectations, this model can no longer keep pace. AI-driven computer vision finally turns one-off audits into a continuous, scalable, evidence-based capability - and Drishti is what that looks like deployed.
Highway authorities, municipal corporations, construction firms, and fleet operators all face the same set of structural challenges in road safety auditing - and Drishti was designed around every one of them.
Survey teams driving long routes with clipboards and cameras generate high labor costs per kilometre audited - and that cost grows linearly with network size.
Days or weeks of lag between inspection and report leave authorities and operators acting on stale data while real defects continue to grow.
Subjective scoring and varying field practices produce audits that aren't directly comparable across regions, auditors, or even seasons.
Critical hazards go unnoticed between scheduled inspections, increasing accident risk on the very segments that should be the highest priority.
Manual processes simply cannot match the growth of national, state, and concessionaire road networks - more kilometres means linear hiring, not better coverage.
Field footage varies in resolution, lighting, weather, and camera quality - naive ML pipelines fail under that variation. The model has to expect it.
Drishti is an AI-driven computer vision platform that turns standard vehicle dashcam or smartphone video into a continuous stream of structured highway intelligence. The system detects, segments, and classifies road safety features along any highway, audits them against IRC and MoRTH standards, and produces evidence-based recommendations - all at real-time inference speeds of 45-60 FPS.
Under the hood, Drishti runs an ensemble of YOLOv8-L for high-accuracy object detection, YOLOv11-L for next-generation detection performance, and a lightweight YOLOv11n classifier for surface classification at scale. Ensemble routing sends easy frames to the small model and hard frames to the large model, trained on 34,540+ labeled images covering 8+ asset categories.
Every detection comes with a confidence score so auditors know which to verify. GPS geo-tagging on every frame means inspectors get a per-kilometre asset inventory, not just a defect list. Compliance reports are formatted for NHAI, state authorities, and concessionaires with audit logging and version control baked in.
A walkthrough of the Drishti road safety AI - from standard dashcam video capture through cloud ingestion and YOLOv8/v11 ensemble inference to the auditor dashboard showing GPS-tagged defects, confidence scores, and IRC/MoRTH/NHAI compliance reports in real time.
From standard vehicle video capture through cloud ingestion and the YOLOv8/v11 ensemble to audit/compliance intelligence and stakeholder-facing applications - cloud-native, horizontally scalable, and validated specifically for Indian highway conditions.
Each model captures part of the road-asset detection problem and is wasteful on the rest. Ensemble routing sends small detections to the lightweight classifier and hard frames to the larger detectors - so accuracy stays high without burning compute on every frame.
High-accuracy object detection for potholes, cracks, signs, crash barriers, lane markings, and delineators. The workhorse for hard frames where confident class labels and tight bounding boxes matter most.
Next-generation YOLO architecture for the toughest detection regimes - low light, motion blur, dense clutter, and partial occlusion. Lifts recall on the hardest 10-20% of frames where v8 alone slips.
Fast surface classifier (concrete, asphalt, gravel, damaged) at 97% accuracy. Runs at high FPS on edge devices so dashcams and smartphones can produce surface labels in real time without backhauling video.
Drishti was designed to move three things at once - highway safety compliance, manual inspection cost, and audit reach - in the same direction. Headline detection metrics: 85%+ pothole accuracy, 97%+ surface classification, 82%+ crack typing, 71%+ kerb assessment.
Talk to the Drishti team to learn how computer vision can transform road safety auditing in your region or network - from a single corridor pilot to a fleet-wide rollout across national, state, and concessionaire highways.