7.7.19. Human Fall Detection

Edge AI solution for real-time human fall detection using accelerometer data on MSPM0G5187 with NPU acceleration.

7.7.19.1. Overview

This example demonstrates a safety-critical Edge AI application that classifies human movement as Activities of Daily Living (ADL) or Fall using accelerometer data. Falls are a leading cause of injury and death in elderly populations and industrial settings. The application provides real-time fall detection directly on a microcontroller without cloud connectivity or bulky external hardware.

Application: Elderly care, worker safety, wearable safety systems, occupancy monitoring

Task Type: Time Series Classification

Data Type: Multivariate (3-axis accelerometer)

Key Achievement: 97.65% accuracy with <1 ms inference latency

7.7.19.2. Device Support

The primary target device is the MSPM0G5187 with accelerometer sensor.

Device

Hardware

Configuration File

MSPM0G5187

MSPM0 with NPU + BMI270 accelerometer (TIDA-010997)

config_MSPM0.yaml

7.7.19.3. System Components

Hardware

  • MSPM0G5187 microcontroller with integrated NPU

  • TIDA-010997 EdgeAI Boosterpack with BMI270 3-axis accelerometer

  • BMI270 specifications: 16-bit resolution, ±16g range, 100-200 Hz sampling rate

Software

  • Code Composer Studio (CCS) 12.x or later

  • MSPM0 SDK 2.11.00 or later

  • TI Edge AI Studio

7.7.19.4. Running the Example

cd tinyml-modelzoo
./run_tinyml_modelzoo.sh examples/fall_detection_classification/config_MSPM0.yaml

7.7.19.5. Dataset Description

This example uses the SisFall dataset, a publicly available dataset of fall and normal movement patterns collected from multiple subjects.

Dataset Source:

Sensor Data:

The original SisFall dataset includes three sensors:

  • ADXL345 — 13-bit triaxial accelerometer

  • ITG3200 — triaxial gyroscope

  • MMA8451Q — triaxial accelerometer

For this example, only ADXL345 accelerometer data is used to match the single-accelerometer BMI270 hardware on TIDA-010997.

Classes (2):

Class

Description

ADL

Activities of Daily Living (walking, sitting, standing, lying, exercising)

Fall

Actual fall events (forward fall, backward fall, lateral fall, etc.)

7.7.19.6. Data Preprocessing

1. Sensor Filtering

Only ADXL345 (triaxial accelerometer) columns are extracted from SisFall:

# Extract only accelerometer (ADXL345)
ax, ay, az = df['ADXL345_x'], df['ADXL345_y'], df['ADXL345_z']
# Drop gyroscope (ITG3200) and other accelerometer (MMA8451Q)

2. Bit-Depth Scaling

ADXL345 is a 13-bit sensor, but BMI270 (deployed hardware) is 16-bit. Scale training data to match deployment:

# Scale from 13-bit to 16-bit resolution
scale_factor = 2**16 / 2**13  # = 8
scaled_value = raw_value * scale_factor

This ensures the training data distribution matches the resolution of live sensor data during inference.

3. Format Conversion

Convert SisFall CSV format to TinyML ModelZoo input format:

  • Parse each recording file

  • Extract 3-axis accelerometer values

  • Label as ADL or Fall

  • Export to standardized training dataset structure

7.7.19.7. Feature Extraction Pipeline

Raw 3-axis accelerometer data → Model input (64 features):

  1. Sensor Input — 3-axis accelerometer (x, y, z) @ 100-200 Hz

  2. Windowing — Segment into fixed-size windows (e.g., 256 samples = 1.3-2.5 sec)

  3. FFT Computation — 256-point Real FFT per axis

  4. Magnitude Calculation — Compute spectral magnitude

  5. DC Removal — Remove DC bias component

  6. Binning — Average 16 adjacent FFT bins → 8 features per axis

  7. Frame Concatenation — Stack 8 frames (8 features × 3 axes × 3 frames = 72, reduced to 64)

  8. Model Input — 64-element feature vector per window

Output: Time-domain and frequency-domain features capture both impulse (fall) and sustained (ADL) patterns.

7.7.19.8. Model Architecture

A generic time-series classification CNN optimized for MCU deployment:

Component

Value

Flash (KB)

RAM (KB)

Notes

Model

CLS_6k (6,000 params)

14

0.7

INT8 quantized

Input Shape

64 features

—

—

FFT + time-domain

Layers

Conv2D → MaxPool → Dense → Dense

—

—

Standard CNN

Quantization

INT8

—

—

Weights + activations

Performance Metrics (on LP-MSPM0G5187 Launchpad):

Metric

Value

Target

Notes

Accuracy

97.65%

≥97.5%

On SisFall test set

Inference Latency

0.67 ms

<5 ms

On NPU

Model Size

14 KB

<50 KB

INT8 quantized

RAM Runtime

0.7 KB

<16 KB

Activations only

7.7.19.9. Configuration

File: config_MSPM0.yaml

common:
    task_type: "timeseries_classification"
    model_name: "generic_timeseries_cnn"

data_processing_feature_extraction:
    variables: 3                    # 3-axis accelerometer
    window_length: 256              # 256 samples per window
    frame_count: 8                  # 8 frames (3.2-6.4 sec @ 100-200 Hz)
    feat_ext_transform:
        - "FFT"
        - "MAGNITUDE"
        - "DC_REMOVAL"
        - "POOLING"
    data_proc_transforms:
        - "SCALING"
        - "NORMALIZATION"

training:
    quantization: 2                 # TI-optimized quantization
    quantization_bit_width: 8       # INT8
    epochs: 100
    batch_size: 32
    learning_rate: 0.001

compilation:
    target_device: "MSPM0G5187"
    compiler: "nnc"
    optimization_level: "o3"

7.7.19.10. Deployment Workflow

1. Real-Time Inference

// MCU firmware loop
while (true) {
    // Read 3-axis accelerometer from BMI270
    read_accelerometer(accel_x, accel_y, accel_z);

    // Apply feature extraction (FFT, DC removal, etc.)
    extract_features(accel_x, accel_y, accel_z, features);

    // Run inference on NPU
    classify(features, &class_probabilities);

    // Check for fall
    if (class_probabilities[FALL] > THRESHOLD) {
        trigger_alert();
    }
}

2. Alert Generation

When fall is detected (confidence > 90%): - Sound alarm - Send alert to caregiver (via Bluetooth or network) - Log event with timestamp - Wait for user cancellation (false fall reset button)

3. Sensor Calibration

For new users/environments: - Collect 30-60 seconds of ADL data - Fine-tune model using On-Device Training (ODT) - Improve accuracy for person-specific characteristics

7.7.19.11. On-Device Training

This example supports on-device training for personalization:

ondevice_training:
    enabled: true
    split_layer: "before_dense"
    trainable_layers: 1
    learning_rate: 0.001
    epochs_per_batch: 5

With ODT enabled: - Device collects local ADL samples during normal use - Fine-tunes classification head to user’s movement patterns - Reduces false positives over first week of deployment - Maintains high accuracy as user ages or activity changes

7.7.19.12. Practical Considerations

1. Sensor Placement

  • Mount accelerometer on belt or wrist

  • Keep sensor level with ground (minimize static tilt)

  • Ensure firm contact to avoid loose vibration

2. Sampling Rate

  • Default: 100-200 Hz

  • Nyquist limit for fall detection: ~50 Hz (capture impulses)

  • Higher rates: better impulse fidelity, more power consumption

3. False Positives / False Negatives

  • False positives (ADL misclassified as fall): Rapid movements (jumping, running) - Mitigate: Collect jumping/running samples, on-device training

  • False negatives (fall misclassified as ADL): Slow falls (controlled descent) - Model trained on natural falls; handled by SisFall dataset diversity

4. Battery Life

  • NPU inference: <10 mW @ 0.67 ms

  • Continuous sampling @ 100 Hz: ~100 mW accelerometer

  • Typical wearable: 24-48 hours on coin cell

5. Privacy

  • All inference happens on-device

  • No data transmission to cloud

  • Raw accelerometer data never leaves device

  • GDPR/HIPAA compliant

7.7.19.13. Cross-References

Related Examples:

Related Features:

7.7.19.14. Troubleshooting

Low accuracy in deployment:

  • Enable on-device training for user-specific adaptation

  • Verify sensor placement and calibration

  • Collect more fall samples if available

Frequent false positives:

  • Collect jumping/running ADL samples

  • Increase fall confidence threshold (e.g., 95% instead of 90%)

  • Fine-tune model with local ADL data

High latency:

  • Verify NPU is enabled in device config

  • Check CPU clock is not throttled

  • Reduce feature dimension if needed

Sensor not reading:

  • Check I2C bus (BMI270 uses 0x68 or 0x69 I2C address)

  • Verify TIDA-010997 firmware is up-to-date

  • Test accelerometer directly in CCS debugger

7.7.19.15. References