Contact person: Business Client

Phone:Show

Email:Show

Location: Hyderabad, India

Budget: Recommended by industry experts

Time to start: As soon as possible

Project description:
"Project: On-device Human Activity Recognition (HAR) for Android.
This project delivers a fully on-device Human Activity Recognition (HAR) system for Android that classifies everyday motions—walking, jogging, sitting, standing, and lying—using only the phone’s inertial sensors. All computation runs locally to preserve privacy and guarantee low latency even without connectivity. The pipeline reads accelerometer and gyroscope streams at 20–50 Hz, forms overlapping windows of 128 samples across six channels, and applies the same z-score normalization computed on the training split to ensure training–inference parity. A compact CNN→LSTM model learns short motion motifs with 1D convolutions and aggregates them temporally with an LSTM before a softmax layer outputs activity probabilities.
To satisfy embedded constraints, the floating-point model trained in Keras/TensorFlow is exported to TensorFlow Lite and quantized to INT8 using post-training quantization with a class-balanced calibration set of at least 500 windows. Quantization typically reduces size by about 4× and speeds inference while keeping accuracy within roughly 0–2 percentage points of the FP32 baseline. The Kotlin Android app integrates the TensorFlow Lite interpreter, prefers NNAPI when available (with CPU fallback), and logs per-inference latency so the evaluation reflects real device performance.
Evaluation follows a strict subject-wise split to avoid identity leakage and reports overall accuracy, macro-F1, model size, and on-device latency (mean ± standard deviation over ≥200 windows). Project targets are ≥90% test accuracy, ≤15 ms per inference, and an INT8 model near or below 1 MB. Deliverables include reproducible preprocessing code, trained FP32 and INT8 models, calibration assets, a live demo app showing predictions and latency, and a concise accuracy–size–latency comparison table suitable for inclusion in the final report." (client-provided description)