ML Object Detection: Pothole Detection (TensorFlow 1)

Overview

This project was developed as part of my Bachelor of Engineering final year project. The system implements a pothole detection pipeline using TensorFlow 1 and OpenCV, leveraging object detection techniques for real-time analysis of road surfaces from video footage. It detects potholes, calculates approximate distance, and provides a WARNING!!! alert if a pothole is in the vehicle’s path.

Note: This repository is provided for reference purposes only. It contains scripts and configurations from a 2019 TensorFlow 1 pothole detection project. The code may require outdated dependencies (Python 3.5, TensorFlow 1.x) and is not intended for direct use.

---

Key Features

• Video Input: Supports webcam stream or pre-recorded video.
• Object Detection: Uses SSD MobileNet v1 pretrained model with custom pothole detection modifications.
• Distance Estimation: Calculates relative distance of potholes to the vehicle path.
• Alerts: Displays WARNING!!! when a pothole is directly in vehicle path.

---

Folder Structure

ml-object-detection-potholes-tf1/
│
├── scripts/ # Pothole detection Python scripts
│ └── object_detection_video.py
├── utils/ # Helper modules from TF Object Detection API
│ ├── label_map_util.py
│ └── visualization_utils.py
├── data/ # Label map and optional demo images
│ ├── labelmap.pbtxt
├── requirements.txt # Required Python packages
└── README.md

---

Requirements

tensorflow==1.15.0
opencv-python==3.4.3.18
numpy==1.16.4
pillow==6.2.1
matplotlib==2.2.3
Cython==0.29.10
contextlib2==0.5.5
lxml==4.2.5
jupyter==1.0.0
protobuf==3.6.1  # optional for .proto compilation

---

Setup & Installation

These steps are included for documentation and reference only. Running them may require legacy Python and TensorFlow versions.

1. Install Python 3.5 (amd64)
2. Create TensorFlow folder: C:\TensorFlow
3. Install TensorFlow 1.x:

pip install tensorflow==1.15.0
pip install numpy==1.16.4 pillow==6.2.1 matplotlib==2.2.3
pip install Cython==0.29.10 contextlib2==0.5.5 lxml==4.2.5
pip install jupyter==1.0.0
pip install opencv-python==3.4.3.18

4. Download SSD MobileNet v1 COCO model from the TensorFlow Model Zoo and extract it into:

5. Clone this repository:

git clone https://github.com/siddhant-savant/ml-object-detection-potholes-tf1.git
cd ml-object-detection-potholes-tf1

in C:\TensorFlow\models\research\object_detection

6. Run the object detection script (webcam or local video):

python scripts\object_detection_video.py

---

Usage

Webcam Input

cap = cv2.VideoCapture(0)

Local Video Input

cap = cv2.VideoCapture('C:/path/to/video.mp4')

Detection

• Bounding boxes highlight potholes.

• Approximate distance displayed above each detected pothole.

• Displays WARNING!!! if pothole is directly in vehicle path.

References

TensorFlow Object Detection API: https://github.com/tensorflow/models/tree/master/research/object_detection

SSD MobileNet v1 Model Zoo: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md

LabelImg annotation tool: https://tzutalin.github.io/labelImg/Usage