Research Intern - Vehicle Indicator Detection System

Published:

Project Overview

This project focused on advancing Intelligent Transportation Systems (ITS) by creating a highly accurate and stable Deep Learning system capable of detecting vehicle turn indicators from challenging real-world video and image data. The core deliverable involved a rigorous comparative analysis of leading object detection and multi-object tracking frameworks to identify the optimal solution for real-time deployment.

Technology Stack

  • Primary Tools: YOLOv12, RT-DeTR, BoT-SORT, ByteTrack
  • Languages: Python
  • Concepts: Deep Learning, Computer Vision, Multi-Object Tracking (MOT)
  • Pre-processing: CLAHE, Bilateral Filtering, Wiener Filtering

Key Achievements & Findings

  • Model Benchmarking: Systematically trained, validated, and comparatively analyzed four cutting-edge object detection models: YOLOv8, YOLOv11, YOLOv12, and RT-DeTR on a custom dataset.
  • Optimal Solution Identified: Identified YOLOv12 as the optimal model, which provided the best balance between detection accuracy and stable bounding box localization crucial for ITS applications.
  • Real-time Tracking: Implemented and evaluated multi-object tracking algorithms (ByteTrack and BoT-SORT) to ensure stable, real-time tracking performance in sequential video analysis.
  • Problem Mitigation: Successfully diagnosed and resolved severe overfitting and validation instability encountered during the initial investigation of a CNN-LSTM model, enabling a successful pivot to superior object detection frameworks.


Technical Deep Dive: Key Skills Demonstrated

The research involved proficiency across the entire computer vision pipeline:

Skill CategoryDemonstrated Expertise
Deep LearningTraining, testing, and rigorous validation of neural networks.
Object DetectionImplementing and comparing state-of-the-art models (YOLOv12 vs. RT-DeTR).
Multi-Object TrackingEvaluating tracking algorithms (ByteTrack and BoT-SORT) for sequential data.
Model EvaluationBenchmarking performance using metrics like Precision, Recall, mAP@50, and mAP@50-95.
Data PreprocessingApplying image enhancement techniques (CLAHE, Bilateral/Wiener Filtering) to improve model robustness.