NOVA - Networked Onboard Vehicle Analytics

NOVA is an advanced AIoT (Artificial Intelligence of Things) vehicle monitoring and safety system developed as the final project for an internship at the National Telecommunication Institute (NTI).

The system transforms a standard vehicle into an intelligent, connected platform by combining real-time sensor data, edge-based machine learning, sensor fusion, and remote telemetry to enhance driver safety, vehicle security, and predictive maintenance.

Features

• Accurate Location Tracking
  GPS + IMU sensor fusion for continuous positioning (even during temporary GPS loss)

• Cabin Environment Monitoring
  Temperature and humidity sensing using DHT22

• Predictive Engine Health Monitoring
  Vibration + engine temperature analysis with on-device ML model (faulty / not faulty classification)

• Driver Safety & Fatigue Detection
  Camera-based drowsiness and distraction detection using edge ML

• Security Features
  Ultrasonic sensor for prolonged presence detection near driver door (anti-theft alert)

• Headlight Status Monitoring
  Detection of front light failure/damage

• Critical Alerting
  Instant SMS notifications via GSM module for high-priority events

• Remote Telemetry
  Real-time data publishing to MQTT broker over Wi-Fi (JSON payloads)

• Real-Time Multitasking
  Efficient task management using FreeRTOS on ESP32

Hardware Components

Component	Interface	Purpose
ESP32 (DevKit / Custom)	-	Main microcontroller
GPS Module	UART	Location tracking
IMU (e.g., MPU-6050)	I2C	Motion data & sensor fusion
DHT22	Digital	Cabin temperature & humidity
Vibration Sensor + Temp	ADC	Engine health input for ML
Ultrasonic Sensor	GPIO	Proximity detection (driver door)
Light Sensor / Current	ADC/GPIO	Headlight status
Camera Module (e.g., OV2640)	SPI/Parallel	Driver state monitoring
SIM Module (e.g., SIM800L)	UART	SMS alerts

System Architecture

flowchart LR
    subgraph External["External Systems"]
        direction TB
        MQTT[MQTT Broker]
        SMS[SMS Gateway]
    end

    subgraph ESP32["ESP32 (FreeRTOS)"]
        direction TB
        ML["ML Inference\n(Engine Fault & Driver State)"]
        Scheduler[Task Scheduler]
        Fusion[Sensor Fusion]
        Actuator[Actuator Control]

        ML --> Scheduler
        Fusion --> Scheduler
        Actuator --> Scheduler
    end

    subgraph Hardware["Hardware Interfaces\n(Sensors & Modules)"]
        direction TB
        GPS["GPS\n(UART)"]
        IMU["IMU\n(I2C)"]
        DHT["DHT22\n(Digital)"]
        Vib["Vibration Sensor\n(ADC)"]
        Temp["Engine Temp\n(ADC)"]
        Ultra["Ultrasonic\n(GPIO)"]
        Light["Light Sensor\n(ADC/GPIO)"]
        Camera["Camera\n(SPI/Parallel)"]
        SIM["SIM Module\n(UART)"]
    end

    External -->|"MQTT over Wi-Fi"| ESP32
    External -->|"UART (AT commands)"| ESP32
    ESP32 -->|"Data Acquisition\n(UART, I2C, ADC, GPIO)"| Hardware
    Hardware -->|"Sensor Data"| ESP32

    style External fill:#f0f8ff,stroke:#333,stroke-width:2px
    style ESP32 fill:#e6f3ff,stroke:#0066cc,stroke-width:3px,color:#000
    style Hardware fill:#fff0f0,stroke:#cc0000,stroke-width:2px

Loading

Machine Learning Models

NOVA incorporates two custom-trained, lightweight machine learning models optimized for on-device inference on the resource-constrained ESP32 microcontroller using TensorFlow Lite for Microcontrollers.

1. Engine Health Monitoring Model

• Objective: Binary classification to detect anomalous (faulty) engine conditions vs. normal operation.
• Input Features:
  • Vibration magnitude and frequency characteristics (extracted from raw ADC readings via FFT or statistical features: RMS, peak-to-peak, variance)
  • Engine temperature (normalized ADC value)
• Training Pipeline:
  • Feature engineering in Python (NumPy/SciPy)
  • Model architecture: Small fully-connected neural network or quantized decision tree
  • Quantization: Full integer quantization (8-bit) to reduce model size (< 50 KB) and inference time
• Output: Fault probability + confidence score; triggers alert if threshold exceeded

2. Driver State Monitoring Model

• Objective: Detect signs of drowsiness or distraction (e.g., eyes closed, yawning, head pose deviation).
• Input: Pre-processed face-region images captured from the in-cabin camera.
• Training Pipeline:
  • On-device face detection and cropping (lightweight Haar cascade or tiny CNN)
  • Classification model: Depthwise separable convolutional neural network (inspired by MobileNetV1 with alpha 0.25) or custom tiny CNN
  • Transfer learning + fine-tuning on collected dataset
  • Quantized to 8-bit integers for edge deployment (model size < 100 KB)
• Output: Drowsy/Distracted probability; triggers escalating alerts (MQTT notification → SMS if persistent)

Both models were trained, validated, and optimized by the NTI internship team to ensure efficient execution within ESP32 memory and power constraints.

Documentation

• Software Requirements Specification (SRS)
• Project Proposal
• System Design Document

All available in the /doc/ and System_Architecturefolder.

License

This project is licensed under the MIT License - see the LICENSE file for details.

---

NOVA – Turning vehicles into intelligent, connected, and safer platforms.