Human Detection System | Ashvath Cheppalli

Key Metrics

89.6%

mAP@50

86.7%

Precision

6-8ms

Inference Speed

92%

Overall Accuracy

Project Overview

Developed an intelligent human detection system using YOLOv8 deep learning model specifically designed for emergency fire evacuation scenarios. The system processes thermal imaging data to detect trapped individuals in low-visibility, smoke-filled environments where traditional camera systems fail.

Fine-tuned YOLOv8 on a custom thermal imaging dataset achieving 89.6% mAP50 and 86.7% precision. Developed a cross-platform Flutter mobile application with integrated video processing pipeline and YOLOv8 inference engine, enabling first responders to analyze thermal footage for post-incident rescue operations with 6-8ms inference speed.

Core Algorithm

1 Upload thermal video through Flutter mobile application interface

2 Segment video into individual frames and resize to 640x640 pixels

3 Process frames through YOLOv8 model with CSPDarknet53 backbone

4 Apply Non-Maximum Suppression to filter redundant detections

5 Generate bounding boxes with class labels and confidence scores

6 Display results with annotated frames showing detected individuals

Technologies Used

YOLOv8 Flutter Python TensorFlow Lite OpenCV Dart Thermal Imaging

Technical Implementation

YOLOv8 Fine-tuning: Trained on custom thermal imaging dataset with data augmentation achieving 89.6% mAP50
Mobile Optimization: TensorFlow Lite quantization for 6-8ms inference speed on mobile devices
Video Processing Pipeline: Frame extraction and resizing to 640x640 for optimal model performance
Cross-platform Deployment: Flutter framework ensuring consistent experience across iOS and Android

Key Features

Thermal Imaging Support: Optimized for low-visibility smoke-filled environments using heat signatures
Real-time Processing: 6-8ms inference enables near real-time detection on mobile hardware
Offline Capability: TFLite model runs entirely on-device without internet connection
Emergency Response Focus: Designed specifically for post-incident rescue analysis scenarios

Challenges & Solutions

Low Visibility Detection: Trained model specifically on thermal imaging data to detect heat signatures in smoke-filled environments
Mobile Performance: Optimized using TensorFlow Lite quantization achieving 6-8ms inference on resource-constrained devices
Dataset Limitations: Created custom thermal imaging dataset with diverse fire scenarios and human positions for robust training

Results & Impact

The system achieved 89.6% mAP50 and 86.7% precision on thermal imaging test dataset, demonstrating high accuracy in detecting humans in low-visibility conditions. The 6-8ms inference speed enables near real-time processing on mobile devices, making it practical for emergency response scenarios.

The Flutter mobile application provides an intuitive interface for first responders to analyze thermal footage from fire scenes, potentially saving lives by quickly identifying trapped individuals in post-incident rescue operations. The system represents a significant advancement in emergency response technology for fire evacuation scenarios.

Intelligent Human Detection System For Emergency Fire Evacuation