Skip to content

fall_detection_classification

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

fall_detection_classification

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
config_MSPM0.yaml
config_MSPM0.yaml
 
 
readme.md
readme.md
 
 

readme.md

Human Fall Detection Classification

Overview

The Human Fall Detection Classification application is an Edge AI solution that classifies human movement into two states: Activities of Daily Living (ADL) or Fall, using accelerometer data in real-time. This enables safety monitoring on embedded devices and provides immediate alerts in belt-mounted safety systems for vulnerable individuals such as elderly users or industrial workers.

Problem and Solution

  • Falls are a leading cause of injury and fatality among elderly individuals and in industrial environments
  • Traditional fall detection systems rely on cloud connectivity or bulky wearable hardware
  • Early and accurate fall detection enables timely emergency response and reduces injury severity
  • Edge AI enables real-time fall classification directly on resource-constrained microcontrollers without cloud connectivity

Key Performance Targets

  • Real-time classification of human movement into ADL or Fall
  • High accuracy on accelerometer-based fall detection (≥97.5%)
  • Low memory footprint suitable for MCU deployment

System Components

1. Hardware:

2. Software:

  • Code Composer Studio 12.x or later
  • MSPM0 SDK 2.11.00 or later
  • TI Edge AI Studio

Dataset Information

The example uses the SisFall dataset as the primary training dataset:

Each file contains readings from three sensors:

  • ADXL345 — 13-bit triaxial accelerometer
  • ITG3200 — triaxial gyroscope
  • MMA8451Q — triaxial accelerometer

Data Cleaning and Preprocessing

The SisFall dataset contains data from three sensors, but only the ADXL345 accelerometer data is used for this example to match the single-accelerometer hardware setup (BMI270 on TIDA-010997). The following cleaning and preprocessing steps are applied:

1. Sensor Filtering:

  • Only ADXL345 (13-bit triaxial accelerometer) columns are retained
  • ITG3200 (gyroscope) and MMA8451Q (accelerometer) columns are discarded

2. Bit-Depth Scaling:

  • The ADXL345 is a 13-bit accelerometer, while the target hardware uses the BMI270 which is a 16-bit accelerometer
  • To match the input resolution expected by the firmware, all ADXL345 readings are scaled up from 13-bit to 16-bit resolution using the formula: scaled_value = raw_value * (2^16 / 2^13) = raw_value * 8
  • This ensures consistency between the training data distribution and the live sensor data fed during inference on the MCU

3. Format Conversion:

  • The cleaned and scaled dataset is converted into a format compatible with tinyml-tensorlab for model training
  • Each file is parsed, labeled (ADL or Fall), and exported into the required input structure

Feature Extraction Pipeline

  1. Sensor Input: 3-axis accelerometer data (x, y, z) from ADXL345 (scaled to 16-bit)
  2. Real FFT: 256-point FFT using ARM CMSIS-DSP
  3. Complex Magnitude Calculation
  4. DC Removal
  5. Binning: Average 16 adjacent FFT bins → 8 features
  6. Frame Concatenation: Stack 8 frames (64 total features per axis)

Model Architecture

A generic time-series classification CNN model is used:

Model Parameters Flash (kB) RAM (kB) Inference Latency (NPU) Accuracy
CLS_6k ~6,000 14 0.7 0.67 ms 97.65%

NOTE: The above statistics was measured on LP-MSPM0G5187 Launchpad

The model uses:

  • INT8 quantization for reduced memory footprint
  • 3-channel input (x, y, z axes)

Training and Deployment Process

NOTE: Running the config yaml handles everything including dataset loading, data cleaning, feature extraction, training, quantization, and compilation.

  1. Training:

    • Use TI Edge AI Studio (GUI) or tinyml-tensorlab (CLI)
    • Batch size: 256
    • Training epochs: 50

  2. Quantization:

    • INT8 quantization for reduced model size
    • Maintains accuracy while enabling MCU deployment

  3. Compilation:

    • TI Neural Network Compiler converts the trained model
    • Generates model artifacts for device deployment

How to Run

After completing the repository setup, run the following command from the tinyml-modelzoo directory:

Windows:

.\run_tinyml_modelzoo.bat examples\fall_detection_classification\config_MSPM0.yaml

Linux:

./run_tinyml_modelzoo.sh examples/fall_detection_classification/config_MSPM0.yaml

References