These are quantized TensorFlow Lite models designed for binary person classification (person vs. non-person) optimized for embedded deployment on microcontrollers, specifically targeting the Astra MCU SDK platform.

The models perform real-time object detection and classification, analyzing camera feed inputs to accurately identify human presence within the field of view.

Model Variants

This repository contains two optimized model variants for different deployment scenarios:

1. Flash Memory Model

File: person_classification_flash(448x640).tflite
Input Resolution: VGA (640×480×3)
Target: Flash memory deployment for higher resolution processing
Use Case: Applications requiring higher accuracy with sufficient flash storage
Memory Requirement: Model stored in flash memory

2. SRAM Model

File: person_classification_sram(256x448).tflite
Input Resolution: WQVGA (480×270×3)
Target: SRAM deployment for memory-constrained environments
Use Case: Real-time applications with limited memory footprint
Memory Requirement: Model stored in SRAM for faster access

Technical Specifications

Input Format

Data Type: uint8 (quantized RGB values 0-255)
Channel Order: RGB
Normalization: No additional normalization required (quantized model)

Flash Model Input Shape: [1, 480, 640, 3] (921,600 total values) SRAM Model Input Shape: [1, 270, 480, 3] (388,800 total values)

Output Format

Task: Binary classification
Output: Single probability score
Classes:
- 0: Non-person
- 1: Person
Data Type: Quantized output (requires dequantization for probability)

Model Optimization

Quantization: 8-bit quantization for embedded deployment
Model Size: ~1.5MB each (highly optimized for MCU deployment)
Target Hardware: ARM Cortex-M55 with Ethos-U55 NPU
Framework: TensorFlow Lite for Microcontrollers

Usage Example

python

import tensorflow as tf
import numpy as np
from PIL import Image

def load_model(model_path):
    """Load TensorFlow Lite model"""
    interpreter = tf.lite.Interpreter(model_path=model_path)
    interpreter.allocate_tensors()
    return interpreter

def preprocess_image(image_path, target_size):
    """Preprocess input image"""
    image = Image.open(image_path).convert('RGB')
    image = image.resize(target_size)
    image_array = np.array(image, dtype=np.uint8)
    return np.expand_dims(image_array, axis=0)  # Add batch dimension

def classify_person(interpreter, image_data):
    """Run inference on the model"""
    input_details = interpreter.get_input_details()
    output_details = interpreter.get_output_details()
    
    interpreter.set_tensor(input_details[0]['index'], image_data)
    interpreter.invoke()
    
    output_data = interpreter.get_tensor(output_details[0]['index'])
    
    # Dequantize output if needed
    scale = output_details[0]['quantization'][0]
    zero_point = output_details[0]['quantization'][1]
    
    if scale != 0:  # Quantized output
        dequantized_output = scale * (output_data.astype(np.float32) - zero_point)
        probability = 1 / (1 + np.exp(-dequantized_output[0]))  # Apply sigmoid
    else:
        probability = output_data[0]
    
    return probability

# Example usage
# For Flash model (VGA resolution)
interpreter_flash = load_model("person_classification_flash(448x640).tflite")
image_data_vga = preprocess_image("test_image.jpg", (640, 480))
result_flash = classify_person(interpreter_flash, image_data_vga)

# For SRAM model (WQVGA resolution)  
interpreter_sram = load_model("person_classification_sram(256x448).tflite")
image_data_wqvga = preprocess_image("test_image.jpg", (480, 270))
result_sram = classify_person(interpreter_sram, image_data_wqvga)

print(f"Flash model prediction: {result_flash:.3f}")
print(f"SRAM model prediction: {result_sram:.3f}")

Performance Characteristics

Memory Requirements

Flash Model: Designed for devices with sufficient flash storage
SRAM Model: Optimized for memory-constrained microcontrollers
Inference Memory: Minimal tensor arena requirements due to quantization

Deployment Targets

Primary Platform: Astra MCU SDK with ARM Cortex-M55 + Ethos-U55 NPU
Compatible: Any TensorFlow Lite Micro compatible microcontroller
Resolution Support: Multiple input resolutions for different use cases

Real-World Applications

Security and surveillance systems
Smart home automation
IoT edge devices
Battery-powered computer vision applications
Real-time person detection in embedded systems

Deployment Guide

For Astra MCU SDK:

Model Selection:
- Use Flash model for VGA resolution applications
- Use SRAM model for memory-constrained or real-time applications
Integration:
- Place model file in your project's model directory
- Configure input resolution in your application
- Set up camera interface for chosen resolution
Memory Configuration:
- Flash model: Configure flash memory allocation
- SRAM model: Ensure sufficient SRAM for model weights
Build Configuration:

bash
1# For WQVGA (SRAM model) 2make cm55_person_classification_defconfig 3 4# For VGA (Flash model) - modify via menuconfig 5make menuconfig 6# Navigate to: COMPONENTS CONFIGURATION → Off Chip Components → Display Resolution 7# Change to: VGA(640x480)

General TensorFlow Lite Deployment:

python

# Install dependencies
# pip install tensorflow numpy pillow

import tensorflow as tf

# Load and run model
interpreter = tf.lite.Interpreter(model_path="person_classification_sram(256x448).tflite")
interpreter.allocate_tensors()

# Get input and output tensors
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

# Run inference
interpreter.set_tensor(input_details[0]['index'], input_data)
interpreter.invoke()
output_data = interpreter.get_tensor(output_details[0]['index'])

Model Selection Guide

Use Case	Model Variant	Resolution	Memory	Best For
High Accuracy	Flash Model	640×480	Flash storage	Security cameras, detailed detection
Real-time Processing	SRAM Model	480×270	SRAM storage	IoT devices, battery-powered applications
Memory Constrained	SRAM Model	480×270	Low memory	Microcontrollers with limited resources
General Purpose	Both	Variable	Flexible	Development and prototyping

Requirements

TensorFlow Lite runtime (2.13.0+)
NumPy for data processing
PIL/Pillow for image handling
Target hardware with TensorFlow Lite Micro support

Citation

If you use these models in your research or applications, please cite:

bibtex

@misc{person_classification_tflite_2024,
  title={Person Classification TensorFlow Lite Models for Embedded Deployment},
  author={Astra MCU SDK Team},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/JahnaviBhansali/person-classification-tflite}
}

License

Licensed under the Apache License 2.0. See LICENSE file for details.

Related Resources

These models are optimized for embedded deployment and provide an excellent starting point for person detection applications on resource-constrained devices.