Harshit Makwana

Computer vision / Case study

Roadway Solutions

I developed road-inspection workflows using fine-tuned YOLO, ByteTrack, MiDaS and SAM 2, with OpenVINO experiments for inference performance.

Harshit Makwana

Concept illustration of a pothole contour, estimated depth surface and segmented roadside regions.
AI-generated concept illustration · not a project screenshot
My role
Detection, segmentation, tracking & inference optimization
Timeline
2024–ongoing
Focus
Computer vision

01 / Context

Combining object detection, surface analysis and road-scene segmentation.

Road inspection involves different visual tasks: locating and counting objects, examining potholes, and distinguishing road dividers, footpaths and vegetation. I worked on combining these capabilities with location-related notifications in a real-time workflow.

02 / Ownership

What I built

  • Fine-tuned YOLO for real-time object detection and segmentation, with ByteTrack for tracking and counting.
  • Combined MiDaS depth estimation with contours from segmented potholes to support pothole analysis.
  • Tested alternative models and worked on inference performance using OpenVINO.
  • Applied SAM 2 to road-scene segmentation for dividers, footpaths and shrubs or trees, and developed real-time location-related notifications.

03 / Engineering

Technical approach

Use different models for distinct visual tasks

I used fine-tuned YOLO for detection and segmentation and ByteTrack to connect detections over time for tracking and counting. For potholes, I combined the segmented object contour with MiDaS depth estimates to support analysis of the detected road surface.

Extend the scene understanding

I applied SAM 2 to segment road dividers, footpaths and vegetation in the road scene. I also tested different models and used OpenVINO in performance-improvement work, alongside the notification logic associated with detected locations.

04 / System view

Workflow at a glance

  1. Road imagery

    Objects, surfaces and roadside context

  2. Vision tasks

    YOLO / ByteTrack, MiDaS and SAM 2

  3. Scene analysis

    Counts, pothole contours/depth and road regions

  4. Notifications

    Location-related inspection information

05 / Delivery

Outcomes

The ongoing project connects complementary computer-vision tasks into a road-inspection workflow. My work spans model adaptation, tracking, pothole analysis, segmentation and inference optimization.

Technologies & methods

  • YOLO
  • ByteTrack
  • MiDaS
  • SAM 2
  • OpenVINO
  • Computer vision

VehiSense Pro

I developed a vehicle-monitoring system with fine-tuned YOLOv8, tracking and event-detection logic, connected to a FastAPI web application through WebSockets.

Image Inpainting Using GANs

I developed a GAN-based image-inpainting model, trained it on artificially masked images and built a Streamlit interface for uploading images and viewing restorations.

Satellite Image Change Detection

Developed a system to detect spatial changes over time using satellite imagery and advanced image processing techniques. The solution combines deep learning models, clustering, and dimensionality reduction for environmental and urban insights.