Vehicle Detection and Tracking System

A comprehensive vehicle detection, tracking, and traffic analysis system that implements two different approaches:

1. DETR (DEtection TRansformer) based implementation
2. Faster R-CNN based implementation
3. RF-DETR with ConvNext Backbone

First and Second implementations use DeepSORT for object tracking and provide similar functionality with different underlying detection architectures.

RF-DETR

⚠️ Note: RF-DETR training was not completed due to hardware limitations. The model takes significantly longer to train compared to DETR and Faster R-CNN, especially when relying on Apple's MPS backend (MacBook M1/M2 GPU support).

When training on an M1 Mac, PyTorch often falls back to CPU for unsupported operations, leading to extremely slow performance. One such warning observed during training:

"UserWarning: The operator 'aten::grid_sampler_2d_backward' is not currently supported on the MPS backend and will fall back to run on the CPU. This may have performance implications. (Triggered internally at /Users/runner/work/pytorch/pytorch/pytorch/aten/src/ATen/mps/MPSFallback.mm:14.)"

Features

• Real-time vehicle detection using DETR, Faster R-CNN, or RF-DETR.
• Multi-object tracking with DeepSORT
• Vehicle counting and classification
• Traffic analysis and statistics
• Visualization of detection and tracking results
• Support for video and image processing
• Traffic flow analysis and reporting

Requirements

• Python 3.8+
• PyTorch >= 2.0.0
• OpenCV >= 4.8.0
• Other dependencies listed in respective requirements.txt files

Installation

1. Clone the repository:

git clone [your-repository-url]
cd Vehicle_countingv2

2. Download DeepSORT:

   • Clone the DeepSORT repository:

   git clone https://github.com/ZQPei/deep_sort_pytorch.git

   • Copy the deep_sort_pytorch folder into both DETR_vehicle_counting and fasterRcnn_deepsort directories
   • Download the DeepSORT weights from here
   • Place the downloaded weights in both:
     • DETR_vehicle_counting/deep_sort_pytorch/deep/checkpoint/
     • fasterRcnn_deepsort/deep_sort_pytorch/deep/checkpoint/

3. Download Model Weights:

   • DETR model weights: Download
   • Faster R-CNN model weights: Download
   • Place the DETR weights in DETR_vehicle_counting/checkpoints/
   • Place the Faster R-CNN weights in fasterRcnn_deepsort/checkpoints/

4. Install dependencies for DETR implementation:

pip install -r DETR_vehicle_counting/requirements.txt

5. Install dependencies for Faster R-CNN implementation:

pip install -r fasterRcnn_deepsort/requirements.txt

Usage

DETR Implementation

1. Run vehicle detection and tracking:

python DETR_vehicle_counting/track1.py --source [video_path]

2. Analyze traffic data:

python DETR_vehicle_counting/analyze_traffic.py --input [results_path]

3. Visualize predictions:

python DETR_vehicle_counting/visualize_preds.py --input [predictions_path]

Faster R-CNN Implementation

1. Run vehicle detection and tracking:

python fasterRcnn_deepsort/track.py --source [video_path]

Model Training

DETR Model

To train the DETR model on your own dataset:

1. Prepare your dataset in COCO format
2. Configure training parameters in train.py
3. Run training:

python DETR_vehicle_counting/train.py

Faster R-CNN Model

The Faster R-CNN implementation uses a pre-trained model. If you want to fine-tune it:

1. Prepare your dataset in the required format
2. Configure training parameters in the respective script
3. Run training with your custom dataset

Configuration

• DeepSORT parameters can be adjusted in the deepsort_config directory of each implementation
• Model parameters can be modified in the respective script files
• Visualization settings can be customized in the visualization scripts

Output

Both implementations generate:

• Real-time vehicle detection and tracking visualizations
• Traffic analysis reports
• Vehicle counting statistics
• JSON files containing detailed tracking data

Sample outputs and results can be found in the Outputs Folder

Choosing Between Implementations

• DETR Implementation: Better for modern architectures and transformer-based approaches
• Faster R-CNN Implementation: Better for traditional CNN-based approaches and potentially faster inference

Acknowledgments

• DETR: Facebook Research
• DeepSORT: Original Implementation
• DeepSORT PyTorch Implementation: ZQPei's Implementation
• Faster R-CNN: Original Paper