https://github.com/sarwanshah/distracted-pedestrian-classification-using-cnns-2020

Created a dataset of 1,300 images of distracted pedestrians and applied augmentation to expand it by 3x- 4x. Used Faster R-CNN to localize pedestrians and extract distracted individuals. Tested MLP, ANN (HOG), CNN, and VGG16 transfer learning, achieving 23%, 28%, 60%, and 62% accuracy, respectively.
https://github.com/sarwanshah/distracted-pedestrian-classification-using-cnns-2020

artificial-intelligence classification computer-vision convolutional-neural-networks deep-learning machine-learning recurrent-neural-networks transfer-learning

Last synced: about 1 month ago
JSON representation

Created a dataset of 1,300 images of distracted pedestrians and applied augmentation to expand it by 3x- 4x. Used Faster R-CNN to localize pedestrians and extract distracted individuals. Tested MLP, ANN (HOG), CNN, and VGG16 transfer learning, achieving 23%, 28%, 60%, and 62% accuracy, respectively.

• Host: GitHub
• URL: https://github.com/sarwanshah/distracted-pedestrian-classification-using-cnns-2020
• Owner: SarwanShah
• License: mit
• Created: 2025-02-08T00:06:39.000Z (8 months ago)
• Default Branch: main
• Last Pushed: 2025-03-04T06:12:48.000Z (7 months ago)
• Last Synced: 2025-08-23T01:58:03.067Z (about 2 months ago)
• Topics: artificial-intelligence, classification, computer-vision, convolutional-neural-networks, deep-learning, machine-learning, recurrent-neural-networks, transfer-learning
• Language: Jupyter Notebook
• Homepage:
• Size: 2.51 MB
• Stars: 1
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# Distracted Pedestrian Behaviour Classification  

# Summary

Created a dataset of 1,300 images of distracted pedestrians and applied augmentation to expand it by 3x- 4x.  Used Faster R-CNN to localize pedestrians and extract distracted individuals.  Tested MLP, ANN (HOG), CNN, and VGG16 transfer learning, achieving 23%, 28%, 60%, and 62% accuracy, respectively.

## Project Overview  

This project was developed under the **Computer Vision** course at **Habib University** during Spring 2020, aimed at detecting and classifying distracted pedestrian behavior to prevent road accidents. The study focuses on existing pedestrian detection techniques and improves upon them using advanced algorithms such as **Faster R-CNN**, **Convolutional Neural Networks (CNNs)**, and **Transfer Learning**. The project further explores alert mechanisms to prevent distracted pedestrians from accidents.  

**REPORT: https://github.com/SarwanShah/HU_2020_Distracted-Pedistrian-Classification/blob/main/Final-Report.pdf**

---

## Project Features  

- **Pedestrian Detection**: Faster R-CNN for real-time pedestrian identification.

- **Behavior Classification**: Neural networks and CNNs classify pedestrians into various categories.

- **Data Augmentation**: Extensive augmentation to balance dataset classes and improve model performance.

- **Performance Optimization**: Tested hyperparameters, optimizers, and various CNN architectures for accuracy.

- **Transfer Learning**: Pre-trained VGG16 network to enhance feature extraction and classification accuracy.

---

## Dataset Overview  

- Collected from sources like the **Caltech**, **Inria**, **PETA**, and **PennPad** datasets.

- Final dataset included 1,300 images categorized into four classes:  

  - **Alert**: Pedestrian looking straight ahead.  

  - **Distracted**: Pedestrian looking sideways or downwards.  

  - **UpMobile**: Phone up to the ear.  

  - **DownMobile**: Phone held at waist level.  

- Augmentation techniques applied: image flipping, grayscale conversion, brightness, and contrast adjustments.

---

## Project Implementation  

### ➤ **Pedestrian Detection with Faster R-CNN**  

- Used Faster R-CNN to detect pedestrians and extract bounding boxes from images.

- Classification algorithms applied to these extracted pedestrian images.

### ➤ **Behavior Classification Approaches**  

1. **Neural Network (NN)**: Baseline classification using a simple neural network.  

2. **HOG + NN**: Classified images using extracted HOG feature vectors.  

3. **CNN**: Custom CNN architecture optimized for the task, achieving a maximum test accuracy of 60%.  

4. **Transfer Learning**: VGG16 with pre-trained weights, yielding the highest test accuracy (62.5%).  

### ➤ **Optimizations and Challenges**  

- Explored various hyperparameters, optimizers (SGD, Adam, Adagrad), and dropout regularization to minimize overfitting.

- Best performance achieved with **SGD optimizer**, **batch size of 16**, and **image size of 32x96**.

---

## Results Summary  

| Algorithm            | Test Accuracy (Standalone Data) | Test Accuracy (Extracted Data) |

|----------------------|----------------------------------|---------------------------------|

| Neural Network       | 23%                             | 22%                             |

| HOG + NN             | 28%                             | 44%                             |

| CNN                  | 60%                             | 33%                             |

| Transfer Learning    | 62.5%                           | 52%                             |

---

## Recommendations & Future Considerations  

- Expand dataset to include more categories, such as pedestrians carrying equipment or wearing headphones.

- Incorporate environmental context, e.g., determining whether a pedestrian is alert by detecting nearby objects like cars.

- Explore additional transfer learning models for improved classification accuracy.

---