# 🏎 Drift Car Tracking & Zone Analysis Model

## 📌 Overview

This project is a computer vision model designed to **track drifting cars and quantify driver performance** using aerial (drone) footage. The system detects and tracks vehicles during tandem runs and measures how they interact with predefined drift zones.

The current implementation is a **proof of concept**, developed specifically for footage from **Evergreen Speedway in Monroe, Washington**.

---

## 🧠 Model Description

This model uses a YOLO-based framework to:

- Detect drift cars in tandem runs  
- Classify vehicles as:
  - `leader`
  - `chaser`
- Classify zones as:
  - `FrontZone`
  - `RearZone`
- Track vehicles across frames  
- Enable downstream analysis of zone interaction and timing  

### Training Details

- Fine-tuned from a pretrained YOLO model  
- Custom dataset manually annotated  
- Two datasets:
  - **Cars:** Bounding boxes for leader and chaser  
  - **Zones:** Segmentation masks for drift zones  

Zone interaction is computed using geometric methods (polygon overlap + time tracking), not learned directly by the model.

---

## 🎯 Intended Use

Designed for:

- Formula Drift-style competitions  
- Grassroots drifting events  
- Experimental motorsports analytics  

### Example Applications

- Measuring time spent in drift zones  
- Analyzing tandem behavior (leader vs. chaser)  
- Supporting judging with quantitative insights  
- Enhancing broadcast overlays  

---

## 📊 Training Data

### 📁 Source

- Formula Drift Seattle 2025 PRO, Round 6 - Top 32  
https://www.youtube.com/watch?v=MuD-uxGQnrg&t=879s  

---

### 🔢 Dataset Size

#### 🚗 Cars
- 1,204 original → 2,724 augmented  
- Split: 84% train / 12% val / 5% test  

#### 🟣 Zones
- 724 original → 1,666 augmented  
- Split: 85% train / 9% val / 6% test  

---

### 🏷 Class Distribution

| Class      | Count |
|-----------|------|
| Leader    | 1,204 |
| Chaser    | 1,201 |
| FrontZone | 137   |
| RearZone  | 588   |

---

### ✏️ Annotation

- Fully manual annotation  
- Consistent labeling across frames  
- Handled occlusion, overlap, and tandem proximity  

---

### 🔧 Augmentation

- Rotation ± 8°
- Saturation ± 15%
- Brightness ± 10%
- Blur 2px
- Mosaic = 0.2
- Scale ± 15%
- Translate ± 5% 
- Hsv_h  (Color tint) =  0.01
  
---
## ⚙️ Training Procedure
- **Framework:** Ultralytics YOLO  
- **Models:**
  - Cars: YOLO26s  
  - Zones: YOLO26s-seg  
### 💻 Hardware
- NVIDIA A100 (Google Colab)
### ⏱ Training Time
- Cars: 80 epochs (~42 min)  
- Zones: 140 epochs (~1h 14min)  
### ⚙️ Settings
- Batch size: 16
- Image size: 1024
- Workers: 8
- Cls: 2.5 (Only for Object Detection)
- No early stopping  
- Default preprocessing  
---
## 📈 Evaluation Results
### 🚗 Car Model
| Metric    | Value  |
|----------|-------|
| Precision | 0.9904 |
| Recall    | 0.9792 |
| mAP@50    | 0.9882 |
| mAP@50-95 | 0.8937 |
### 🟣 Zone Model
| Metric    | Value  |
|----------|-------|
| Precision | 0.9919 |
| Recall    | 0.9952 |
| mAP@50    | 0.9948 |
| mAP@50-95 | 0.7064 |
---
## 📉 Key Visualizations
**Car Results**
<img src='results_cars.png' width="800">
<img src='confusion_matrix_cars.png' width="800">
**Zone Results**
<img src='results_zone.png' width="800">
<img src='confusion_matrix_zone.png' width="800">
---
## 🧠 Performance Analysis
### 🚗 Cars
**Strengths:**
- Very high precision and recall  
- Reliable detection and classification  
- Strong tracking foundation  
**Limitations:**
- Smoke occlusion affects detection  
- Close tandem overlap can cause confusion  
- Limited generalization beyond training conditions  
---
### 🟣 Zones
- High detection accuracy (mAP@50)  
- Lower boundary precision (mAP@50-95)
**Implication:**
- Good at identifying zones  
- Less accurate for exact boundaries → impacts timing precision  
**Note:** Since zones are static, polygon-based methods may be more reliable than segmentation.
---
## ⚠️ Limitations and Biases
### 🚨 Failure Cases
- Heavy smoke → missed or unstable detections  
- Close tandem → overlap confusion  
- Camera motion → inconsistent zone alignment  
- Edge-of-frame → partial detections  
---
### 📉 Weak Areas
- Zone boundary precision  
- Leader vs. chaser ambiguity in tight proximity  
---
### 📊 Data Bias
- Single track (Evergreen Speedway)  
- Single event and lighting condition  
- Fixed drone perspective  
---
### 🌦 Environmental Limits
Performance may degrade with:
- Smoke, blur, or occlusion  
- Lighting changes  
- Drone altitude variation  
- Camera movement  
---
### 🚫 Not Suitable For
- Official judging systems  
- General vehicle detection  
- Different tracks without recalibration  
- Other motorsports without adaptation  
---
### 📏 Dataset Limitations
- Underrepresented zone classes  
- Limited diversity (track, cars, conditions)  
- Few edge-case scenarios (spins, collisions)  
---
## 🏁 Summary
This model performs strongly within a controlled environment but is highly specialized. It should be viewed as a **proof-of-concept system** for drift analytics rather than a fully generalized or production-ready solution.