📘 Third Year Project – Road Bend Classification

Project Report: Overleaf Link

Presentation: Video

Presentation Slides: Reveal.js

Project Log: Notion Link

---

🎯 Overview

This project aims to enhance Advanced Driver Assistance Systems (ADAS) by predicting the severity and direction of upcoming road bends using dashcam footage. It introduces a bespoke UK road dataset and proposes a deep learning-based classification system that integrates optical flow and RGB representations from stereo dashcam video.

---

💡 Ideas Explored

This research investigates the effectiveness of human-inspired perception cues in machine-based road understanding. Specifically:

• 🧠 H1: Deep Neural Networks (DNNs) can generalise the task of road bend severity and direction classification effectively from video input.
• 👁️ H2: Optical flow representations, inspired by human visual motion processing, improve classification performance compared to static RGB imagery.
• 🎯 H3: Focused regions around the Road Vanishing Point (R-VP), emulating human gaze, result in better or more efficient classification performance.

Each investigation is evaluated using separate datasets (RGB vs. Optical Flow, Wide vs. Narrow views) and a consistent (2+1)D CNN architecture.

---

📊 Dataset & Resources

Resource	Link
Final Dataset	UK-Road-Bend-Classification
Trained Models	RGB & Optical Flow Models
Raw Dashcam Videos	UK-Road-DashCam
Component Testing Dataset	Stereo-Road-Curvature-Dashcam
Source Code	GitHub Repo
Calibration Files	Camera Calibration

---

📈 Key Results

• RGB Wide View outperformed other configurations, achieving 73.78% accuracy.
• Optical Flow Narrow View had 55.55% accuracy, supporting human-based motion field inspired representation.
• Extensive evaluation with F1 scores and confusion matrices confirms the efficacy of the models under varied driving conditions.

---

🔍 Future Work

• Integrate stereo depth estimation to improve bend understanding.
• Use microcontroller-driven ego-motion correction.
• Experiment with transformer-based or hybrid DNN architectures.

---

⚙️ Processing Pipeline

The processing pipeline is composed of several modular stages:

1. Bend Detection & Labelling

   • Based on GPS heading change over distance-normalised segments.
   • Uses NMEA-formatted GPS embedded in video.

2. R-VP Estimation

   • Road Vanishing Point detection using optical flow and feature tracking.

3. Optical Flow Analysis

   • Dense optical flow and rotational homography correction are applied to capture motion cues.

4. Dataset Generation

   • Wide and narrow field-of-view inputs (emulating driver gaze).
   • Balanced class distributions using SMOTE and hybrid sampling.

5. Deep Learning Classification

   • Utilises (2+1)D CNN to model spatio-temporal features.
   • Compares performance between RGB and optical flow inputs in wide vs. narrow views.

---