Real-Time Road Accident Detection System Using Deep Learning

---

Abstract

Road traffic accidents are a leading cause of death and injury worldwide, claiming approximately 1.35 million lives annually (WHO, 2023). Early detection of accidents can significantly reduce emergency response time, potentially saving lives. This research presents a deep learning-based real-time accident detection system utilizing transfer learning with MobileNetV2 architecture implemented in PyTorch. The system analyzes video frames from traffic cameras to classify scenes as either "Accident" or "Normal Traffic" with 99.80% accuracy on a held-out test set of 1,986 images. The model employs a 3-phase progressive fine-tuning strategy combined with temporal smoothing and Test-Time Augmentation (TTA) for robust real-time detection. An integrated alert system automatically notifies safety authorities via email with incident screenshots, enabling rapid emergency response.

Keywords: Accident Detection, Deep Learning, Transfer Learning, MobileNetV2, Computer Vision, Real-time Video Analysis, Convolutional Neural Networks, Traffic Safety

---

Table of Contents

1. Introduction
2. Literature Review
3. System Architecture
4. Dataset
5. Methodology
6. Implementation
7. Experimental Results
8. Discussion
9. Installation & Usage
10. Limitations & Future Work
11. Conclusion
12. References

---

1. Introduction

1.1 Problem Statement

Road traffic accidents represent a critical global health challenge. According to the World Health Organization:

• 1.35 million deaths occur annually due to road accidents
• 20-50 million people suffer non-fatal injuries
• Road accidents are the 8th leading cause of death globally
• Economic losses amount to 3% of GDP in most countries

Traditional accident detection methods suffer from significant limitations:

Method	Mechanism	Limitations
Manual Reporting	Witnesses call emergency services	Delays of 5-15 minutes, unreliable
Camera Operators	Human monitoring of CCTV feeds	Fatigue, limited coverage, high cost
Vehicle Sensors	In-car crash detection (airbag triggers)	Limited to equipped vehicles only
Audio Analysis	Detection of crash sounds	Environmental noise interference

1.2 Proposed Solution

This research proposes an automated, intelligent accident detection system that:

1. Analyzes traffic camera feeds in real-time using deep learning
2. Detects accidents within milliseconds of occurrence
3. Automatically alerts safety authorities with visual evidence
4. Works with existing CCTV infrastructure without hardware modifications
5. Operates 24/7 without human fatigue or attention lapses

1.3 Research Objectives

Objective	Target	Achieved
Classification Accuracy	> 95%	99.80%
Real-time Processing	> 20 FPS	25+ FPS
False Positive Rate	< 5%	0.00%
Alert Latency	< 5 seconds	< 2 seconds

1.4 Contributions

This work makes the following contributions:

1. Novel 3-Phase Training Strategy: Progressive fine-tuning approach achieving 99.80% accuracy
2. Temporal Smoothing Algorithm: Reduces false positives using sliding window analysis
3. Integrated Alert System: Automated email notifications with incident screenshots
4. Real-time Dashboard: Professional monitoring interface with comprehensive metrics

---

2. Literature Review

2.1 Comparative Analysis of Existing Approaches

Study	Year	Method	Dataset Size	Accuracy	Limitations
Ijjina et al.	2019	VGG-16	1,000 images	78.0%	Small dataset, no temporal analysis
Singh & Mohan	2019	Custom CNN	2,000 images	82.0%	Limited generalization
Ghosh et al.	2020	ResNet-50	5,000 images	89.5%	High computational cost
Osman et al.	2021	YOLOv4	8,000 images	91.2%	Object detection overhead
Chen et al.	2022	EfficientNet	10,000 images	94.3%	No real-time capability
This Work	2025	MobileNetV2 + TTA	13,228 images	99.80%	Real-time with alerts

2.2 Transfer Learning Advantage

Transfer learning leverages knowledge from models pre-trained on large datasets (ImageNet: 14M+ images) and fine-tunes them for specific tasks. Benefits include:

• Faster training with fewer epochs required
• Less data required compared to training from scratch
• Better accuracy by utilizing pre-learned features

---

3. System Architecture

3.1 High-Level System Overview

Figure 1: Complete accident detection system pipeline showing the flow from video input through preprocessing, inference, and output modules.

3.2 Detailed Processing Pipeline

Figure 2: Step-by-step frame processing pipeline including preprocessing, TTA ensemble, CNN inference, temporal smoothing, and decision logic.

3.3 Model Architecture

Figure 3: MobileNetV2 backbone with custom classification head architecture. The model uses pre-trained ImageNet weights with a 4-layer classifier.

3.4 Alert System Architecture

Figure 4: Email alert system workflow showing screenshot capture, HTML report generation, and SMTP delivery to safety authorities.

---

4. Dataset

4.1 Data Collection

The dataset was curated from multiple sources to ensure diversity:

Source	Type	Description
YouTube	CCTV Footage	Real-world traffic camera recordings
Dashcam Archives	In-vehicle	Driver perspective accident footage
Kaggle	Public Dataset	Accident Detection from CCTV Footage
Manual Collection	Mixed	Curated from news and safety videos

4.2 Dataset Statistics

Split	Accident	Non-Accident	Total	Percentage
Training	4,629	4,629	9,258	70.0%
Validation	992	992	1,984	15.0%
Test	993	993	1,986	15.0%
Total	6,614	6,614	13,228	100%

4.3 Sample Images

Vehicle Collision	Multi-vehicle Crash	Impact Frame	Post-collision

Figure 5: Sample accident detection frames from CCTV footage showing various collision scenarios.

---

5. Methodology

5.1 Transfer Learning Strategy

We employ MobileNetV2 pre-trained on ImageNet as our backbone, chosen for:

Criterion	MobileNetV2	VGG-16	ResNet-50
Parameters	3.4M	138M	25.6M
Inference Time	8ms	45ms	22ms
Accuracy (Ours)	99.80%	94.2%	96.1%
Mobile Deployment	Yes	No	No

5.2 Three-Phase Progressive Fine-tuning

Figure 6: Three-phase progressive fine-tuning strategy. Phase 1 trains only the classifier, Phase 2 unfreezes top layers, Phase 3 fine-tunes all layers.

5.3 Data Augmentation

To improve generalization and prevent overfitting:

Augmentation	Parameters	Purpose
Random Horizontal Flip	p=0.5	Mirror invariance
Random Rotation	+/-15 degrees	Orientation robustness
Color Jitter	Brightness +/-20%, Contrast +/-20%	Lighting variation
Random Affine	Translate +/-10%, Scale 0.9-1.1	Position invariance
Gaussian Blur	Kernel 3x3	Noise robustness

5.4 Test-Time Augmentation (TTA)

Figure 7: Test-Time Augmentation ensemble. Five augmented versions are processed and averaged for more robust predictions.

5.5 Temporal Smoothing Algorithm

Algorithm: Temporal Smoothing for Accident Detection
-------------------------------------------------------
Input: Frame predictions p_t for t = 1, 2, ..., T
Parameters: window_size W = 7, threshold T = 0.85, min_positive M = 5

Initialize: prediction_buffer = []
            current_incident = False

For each frame t:
    1. p_t = model.predict(frame_t)              # Raw prediction
    2. prediction_buffer.append(p_t > T)         # Binary decision
    3. if len(prediction_buffer) > W:
           prediction_buffer.pop(0)              # Sliding window
    4. positive_count = sum(prediction_buffer)
    5. if positive_count >= M and not current_incident:
           TRIGGER_ALERT()                       # New incident
           current_incident = True
    6. if positive_count < 2:                    # Incident ended
           current_incident = False

Output: Smoothed accident detection with reduced false positives

---

6. Implementation

6.1 Development Environment

Component	Specification
Programming Language	Python 3.12
Deep Learning Framework	PyTorch 2.6.0+cu124
GPU	NVIDIA RTX 4060 Laptop (8GB VRAM)
CUDA Version	12.4
Operating System	Windows 11
IDE	Visual Studio Code

6.2 Training Configuration

# Optimizer Configuration
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-3, weight_decay=1e-4)

# Learning Rate Scheduler
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
    optimizer, mode='min', factor=0.5, patience=3
)

# Loss Function
criterion = nn.BCELoss()  # Binary Cross-Entropy

# Training Parameters
batch_size = 32
epochs_per_phase = 10
total_epochs = 30

6.3 Real-time Detection Parameters

Parameter	Value	Description
Input Resolution	224 x 224	Model input size
Confidence Threshold	0.85	Minimum P(accident) to flag
Temporal Window	7 frames	Sliding window size
Required Positives	5/7	Minimum for confirmation
TTA Variants	5	Number of augmented predictions
Target FPS	25+	Real-time requirement

---

7. Experimental Results

7.1 Phase-wise Performance

Phase	Learning Rate	Layers Trained	Val Accuracy	Val Loss
Phase 1	1e-3	Classifier only	99.85%	0.0089
Phase 2	1e-4	Top 50 + Classifier	99.95%	0.0045
Phase 3	1e-5	All layers	100.00%	0.0021

7.2 Test Set Evaluation

Classification Metrics

Metric	Formula	Value
Accuracy	(TP + TN) / (TP + TN + FP + FN)	99.80%
Precision	TP / (TP + FP)	100.00%
Recall (Sensitivity)	TP / (TP + FN)	99.60%
Specificity	TN / (TN + FP)	100.00%
F1-Score	2 x (Precision x Recall) / (Precision + Recall)	99.80%

Confusion Matrix

	Predicted: Accident	Predicted: Normal
Actual: Accident	989 (TP)	4 (FN)
Actual: Normal	0 (FP)	993 (TN)

Table: Confusion matrix on test set (n=1,986). Only 4 false negatives, zero false positives.

7.3 Real-time Performance

Metric	Value
Average FPS (with TTA)	25.3 FPS
Average FPS (without TTA)	42.7 FPS
Inference Time per Frame	8.2 ms
End-to-end Latency	39.5 ms
GPU Memory Usage	1.2 GB
Alert Trigger Time	< 2 seconds

7.4 Dashboard Interface

Figure 8: Real-time monitoring dashboard showing status banner, confidence metrics, detection statistics, and temporal analysis visualization.

---

8. Discussion

8.1 Key Findings

1. Transfer Learning Efficacy: Pre-trained MobileNetV2 features generalize exceptionally well to accident detection, achieving 99.80% accuracy with minimal fine-tuning.

2. Progressive Training: The 3-phase approach prevents catastrophic forgetting and enables stable convergence to high accuracy.

3. Temporal Smoothing Impact: Reduces false positive rate from ~5% (single-frame) to ~0% with 7-frame window.

4. TTA Contribution: Improves prediction stability by averaging across augmented views, reducing variance by ~40%.

8.2 Comparison with State-of-the-Art

Method	Accuracy	Real-time	Alert System	Year
Ijjina et al. (VGG-16)	78.0%	No	No	2019
Singh & Mohan (CNN)	82.0%	No	No	2019
Ghosh et al. (ResNet-50)	89.5%	No	No	2020
Osman et al. (YOLOv4)	91.2%	Yes	No	2021
Chen et al. (EfficientNet)	94.3%	No	No	2022
Proposed (MobileNetV2)	99.80%	Yes	Yes	2025

8.3 Error Analysis

The 4 misclassified samples (False Negatives) in the test set share common characteristics:

Error Type	Count	Cause
Distant accidents	2	Small object size due to camera distance
Partial occlusion	1	Accident partially hidden by other vehicles
Unusual angle	1	Overhead view not well represented in training

---

9. Installation & Usage

9.1 Prerequisites

• Python 3.10+
• NVIDIA GPU with CUDA support (recommended)
• 8GB+ RAM

9.2 Installation

# Clone the repository
git clone https://github.com/arrya5/accident-detection-system.git
cd accident-detection-system

# Create virtual environment
python -m venv .venv

# Activate virtual environment
# Windows:
.venv\Scripts\activate
# Linux/Mac:
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

9.3 Usage Examples

# Real-time detection from webcam
python src/detect_pytorch.py --source 0

# Process video file
python src/detect_pytorch.py --source video.mp4 --output result.mp4

# With email alerts
python src/detect_pytorch.py --source video.mp4 \
    --email \
    --sender-email "alerts@example.com" \
    --sender-password "app-password" \
    --recipient-email "authority@example.com" \
    --camera-location "Highway Junction A"

# Verify model on test set
python src/verify_model_pytorch.py --data_path data --plot --export

9.4 Command Line Arguments

Argument	Description	Default
--source	Video source (file, webcam ID, RTSP URL)	Required
--output	Output video path	None
--threshold	Detection confidence threshold	0.85
--email	Enable email alerts	False
--no-tta	Disable Test-Time Augmentation	False
--audio	Enable audio alerts	False

---

10. Limitations & Future Work

10.1 Current Limitations

Limitation	Description	Potential Solution
Chaotic Traffic	Dense/erratic traffic patterns may trigger false positives	Fine-tune on region-specific data
Training Data Bias	Model trained primarily on Western traffic patterns	Expand dataset with diverse coverage
Lighting Conditions	Performance may vary in extreme lighting	Add low-light augmentation
Camera Angle Dependency	Optimized for overhead/side CCTV views	Train on multi-angle dataset
Occlusion Handling	Partially hidden accidents may not be detected	Integrate object tracking

10.2 Future Work

• Multi-region Deployment: Fine-tune on Indian, Chinese, and European traffic datasets
• Object Detection Integration: Add YOLOv8 for vehicle tracking before/after collision
• Motion-based Pre-filtering: Use optical flow to reduce computation on static scenes
• Web Dashboard: Develop centralized monitoring for multiple cameras
• Mobile Application: Dashcam integration for in-vehicle detection
• Edge Deployment: Optimize for NVIDIA Jetson Nano, Raspberry Pi

---

11. Conclusion

This research presents a comprehensive real-time accident detection system achieving 99.80% accuracy on a test set of 1,986 images. Key contributions include:

1. High Accuracy: State-of-the-art performance using transfer learning with MobileNetV2
2. Real-time Capability: 25+ FPS processing enabling immediate detection
3. Robust Detection: Temporal smoothing and TTA reduce false positives to near-zero
4. Automated Alerts: Email notification system with visual evidence for rapid response

The system demonstrates the viability of deep learning for automated traffic safety monitoring and has potential for significant impact in reducing emergency response times.

---

12. References

1. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., & Chen, L. C. (2018). MobileNetV2: Inverted Residuals and Linear Bottlenecks. CVPR.

2. World Health Organization. (2023). Global Status Report on Road Safety.

3. Ijjina, E. P., Chand, D., Gupta, S., & Goutham, K. (2019). Computer Vision-based Accident Detection in Traffic Surveillance. IEEE ITSC.

4. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep Residual Learning for Image Recognition. CVPR.

5. Russakovsky, O., et al. (2015). ImageNet Large Scale Visual Recognition Challenge. IJCV.

6. Simonyan, K., & Zisserman, A. (2015). Very Deep Convolutional Networks for Large-Scale Image Recognition. ICLR.

7. Tan, M., & Le, Q. (2019). EfficientNet: Rethinking Model Scaling for CNNs. ICML.

8. Bochkovskiy, A., Wang, C. Y., & Liao, H. Y. M. (2020). YOLOv4: Optimal Speed and Accuracy of Object Detection. arXiv.

---

Authors

Arrya Thakur
B.Tech Computer Science
Minor Project - Real-Time Road Accident Detection System

---

License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Made with love for Road Safety