[Python] How to Properly Implement Sliding Windows for Real-Time Activity Recognition from...

We're developing a real-time fall detection system using a wearable sensor (accelerometer and gyroscope) that sends continuous data to Firebase. Our machine learning model was trained on 20-second data windows collected at 80 Hz (total 1600 samples per window). The main challenge is correctly implementing the sliding window logic to continuously collect sensor data and make real-time predictions.

Our current approach involves continuously polling Firebase for new batches of sensor data, normalizing them, and buffering these samples in a deque with a maximum size of 1600. Once the buffer reaches 1600 samples, we run a prediction and reset the buffer accordingly.

However, the predictions are inaccurate, and we suspect our approach to sliding windows or buffering might be flawed.

Here's a simplified snippet illustrating the relevant part of our implementation:

BUFFER_SIZE = 1600 # 20 sec window at 80 Hz
STEP_INTERVAL = 5 # Predict every 5 seconds
data_buffer = deque(maxlen=BUFFER_SIZE)
last_prediction_time = time.time()

while True:
samples = fetch_from_firebase() # fetches new data batch (80 samples per second)

for sample in samples:
# Normalize and append each sensor sample to buffer
reading = np.array([
sample['ax'], sample['ay'], sample['az'],
sample['gx'], sample['gy'], sample['gz']
]) / 10000.0
data_buffer.append(reading)

# Prediction logic (every STEP_INTERVAL seconds, after buffer fills)
current_time = time.time()
if len(data_buffer) == BUFFER_SIZE and (current_time - last_prediction_time) >= STEP_INTERVAL:
window = np.array(data_buffer).reshape(1, -1)
window_scaled = scaler.transform(window)
prediction = model.predict(window_scaled)[0]

last_prediction_time = current_time
handle_prediction(prediction)

• Is this the correct way to implement continuous sliding windows for real-time sensor data?
• Could our current method of handling the buffer or timing intervals be causing prediction inaccuracies?
• Are there better practices or alternative strategies for buffering
  real-time sensor data effectively for predictions?

Continue reading...