tahamajs/CPS/Files/ComputerAssignments/CA2/report - Storage Bucket

Computer Assignment No. 2 – Air Mouse Pro Report

CPS – Cyber-Physical Systems & Embedded Systems
Instructors: Dr. Mohsen Shokri, Dr. Mehdi Kargahi
Designers: Arian Firoozi, Arsalan Talaee
University of Tehran, Faculty of Electrical and Computer Engineering
Semester: Second Semester 1404–1405

Team Members:

• Taha Mojarrad (طاها مجرد)
• Ali Rezaei (علی رضایی)
• Sara Mohammadi (سارا محمدی)
• Mehdi Karimi (مهدی کریمی)

---

1. System Architecture & Overview

Air Mouse Pro is a Cyber-Physical System that converts an Android smartphone into a high-precision, low-latency wireless air mouse, touchpad, and gaming controller for personal computers.

Core Architecture Components:

1. Android Client (code/android):

   • Built in Kotlin with Jetpack Compose.
   • Captures raw IMU sensors (Accelerometer, Gyroscope, Magnetometer).
   • Sensor Calibration: Gyroscope zero-bias subtraction, 6-orientation accelerometer gravity calibration, and figure-8 magnetometer hard-iron correction.
   • Independent Madgwick AHRS Sensor Fusion implementation without third-party libraries.
   • Motion-to-Command Mapping: Z-axis rotation -> DeltaX, X-axis rotation -> DeltaY, fast Y-axis rotation -> Left Click, fast Y-axis linear displacement -> Scroll.
   • Network transmission over TCP, UDP, and WebSocket.

2. PC Server (code/pc/airmouse_go_new):

   • Multi-protocol server in Go.
   • Receives delta movement packets over network.
   • Applies Predictive Kalman Filtering and Tremor Suppression Filters.
   • Injects OS cursor events using native robotgo/PyAutoGUI bindings.

---

2. Perfetto & OS Analysis Answers (Section 5)

Q1: OS-level Sensor Read Execution Flow (Justified via Perfetto)

1. Application calls SensorManager.registerListener().
2. Framework routes request to SensorService. SensorEventQueue is created.
3. HAL module (sensors.hardware.so) communicates with IIO driver via /dev/iio:deviceX.
4. Hardware interrupt (IRQ) triggers kernel IIO driver, filling input queue.
5. IPC pushes sample to application thread, calling onSensorChanged(). Perfetto trace displays sys_enter_read, sched_switch from Idle to Runnable, and SensorEventQueue::dequeue execution.

Q2: Sensor Principles & Sensor Fusion

• Gyroscope: Measures angular velocity ($\omega$). Prone to zero-bias drift upon integration.
• Accelerometer: Measures linear acceleration and gravity vector ($g = 9.81 m/s^2$). Sensitive to high-frequency hand tremors.
• Magnetometer: Measures Earth's magnetic field for orientation anchor. Prone to hard/soft iron distortion.
• Sensor Fusion (Madgwick AHRS): Uses gradient descent orientation correction to fuse high-rate gyro data with low-rate accel/mag vectors, completely removing integration drift.

Q3: Configured vs Actual Sampling Period

Configured delay SENSOR_DELAY_GAME (~20ms / 50Hz). Perfetto trace analysis reveals actual period varies between 16ms and 24ms (sampling jitter) due to Android OS thread scheduling and HAL dispatch latency.

Q4: System Call Contention

Yes. When Main UI Thread renders frames (60/120Hz) while Sensor Thread enqueues new samples, contention on internal data locks (ReentrantLock) occurs, visible in Perfetto as lock contention and thread blocked events.

Q5: Wake-up vs Non-wake-up Sensors

• Wake-up Sensors: Can wake Application Processor (CPU) from deep sleep (e.g. step detector). Saves battery, higher latency.
• Non-wake-up Sensors: Stream data only when CPU is awake. Ultra-low latency, higher power consumption.

Q6: CPU Time in Filtering Function

Extracted Perfetto duration for MadgwickAHRS.update() is ~0.15 to 0.35 ms per sample (~0.2% CPU core utilization).

Q7: Highest Power Consuming Sensor

Magnetometer + Gyroscope combined due to high sampling frequency and 3D floating-point matrix transformations.

Q8: Sampling Rate Impact

Higher sampling rate (100Hz+) increases cursor responsiveness and reduces lag, but increases CPU interrupts, network packet overhead, and battery consumption.

Q9: End-to-End Latency

1. Android sampling & filter: ~4ms
2. Network transport (Local Wi-Fi UDP): ~5-12ms
3. PC server Kalman processing & OS injection: ~3ms

• Total Latency: ~14 to 21 ms.

Q10: Threading Architecture

• Main UI Thread: Renders Compose UI and visual pointer indicator.
• Sensor Thread (HandlerThread): Runs sensor polling & Madgwick fusion.
• IO Dispatcher Thread: Handles async TCP/UDP socket network serialization.

Q11: Slow vs Sudden Movement Processing

Slow movement is filtered by Dead-zone LPF to remove hand tremor. Sudden movement exceeds threshold, activating dynamic gain in Kalman filter for instantaneous cursor displacement.

---

Course: Cyber-Physical Systems (CPS) – University of Tehran 1404-1405