Driving Behavior Classifier

An end-to-end machine learning pipeline that classifies driving behaviors from IMU sensor data (gyroscope + accelerometer) using a 1D Convolutional Neural Network, optimized for deployment on ESP32-S3 microcontrollers.

TLDR

The system identifies 9 driving events in real time from a 6-axis inertial measurement unit (MPU6050):

Class	Driving Event
0	Acceleration
1	Aggressive Accelerate
2	Aggressive Brake
3	Aggressive Left
4	Aggressive Right
5	Brake
6	Idling
7	Left
8	Right

Raw sensor streams are windowed, normalized, and fed into a lightweight 1D CNN (~5,108 parameters) that runs as a quantized INT8 model on a ESP32-S3 (FREENOVE)

The point of this is to embed the quantized model to a small ESP32 so that it can classify truck driving behavior using edge compute.

• The main motivation is to reduce power consumption and latency when classifying driving behavior using IoT
• Compared to sending raw sensor data to the cloud -- which is more resource intensive -- edge compute helps reduce this power draw.

---

Architecture

ML Pipeline

driving_events.db
      │
      ▼
┌─────────────────┐
│ Session Parsing  │  Split by time gaps > 5s within DriverID/TaskID pairs
└────────┬────────┘
         ▼
┌─────────────────┐
│ Sliding Window   │  7 timesteps, hop 3 → overlapping windows (data augmentation)
└────────┬────────┘
         ▼
┌─────────────────┐
│ Z-Score Norm     │  Per-channel mean/std fitted on training set only
└────────┬────────┘
         ▼
┌─────────────────┐
│ Augmentation     │  Gaussian noise (σ=0.05), scaling jitter (0.9–1.1), DC offset
└────────┬────────┘
         ▼
┌─────────────────┐
│ 1D CNN           │  2× Conv1D → MaxPool → Dense → Softmax (4 classes)
└────────┬────────┘
         ▼
┌─────────────────┐
│ Export           │  Keras → TFLite F32 → TFLite INT8 → C header (.h)
└─────────────────┘

Model Architecture

Input (7, 6)                         # 7 timesteps × 6 sensor channels
    │
    ├── Conv1D(16, kernel=3, same) + ReLU
    ├── Conv1D(32, kernel=3, same) + ReLU
    ├── MaxPool1D(2)                 # (7,32) → (3,32)
    ├── Dropout(0.3)
    ├── Flatten
    ├── Dense(32) + ReLU
    ├── Dropout(0.4)
    └── Dense(4) + Softmax          

Design:

• 16/32 filters (not 64/128) — keeps model under 10 KB for ESP32-S3 (512 KB SRAM)
• kernel_size=3 — receptive field covers 3 seconds at 1 Hz
• Single MaxPool(2) — (7,32) → (3,32); sequence too short to pool further
• Dropout 0.3/0.4 — aggressive regularization for a small dataset
• 7-step window — matches the shortest driving event (Sudden Brake)

Training Strategy

• Split: Stratified Group K-Fold by DriverID — prevents data leakage across drivers
• Class weights: Balanced, computed from training distribution
• Early stopping: Patience of 25 epochs (max 200)
• Optimizer: Adam (lr=1e-3) with learning rate reduction on plateau
• Batch size: 16

Deployment Target

Artifact	Size	Purpose
driving_cnn.keras	2.89 MB	Full-precision Keras model
driving_cnn_f32.tflite	953 KB	TFLite float32
driving_cnn_int8.tflite	248 KB	INT8 quantized for ESP32-S3
driving_cnn_model.h	1.55 MB	C header for firmware embedding
norm_params.npy	—	Normalization params for on-device preprocessing

---

Dataset Overview

• Sources:
  • Driving Behavior Dataset — MPU6050 (Yuksel, 2021)
  • Driving Events — smartphone sensors (Goh, 2021)
• Shape: 1,114 rows × 12 columns
• Features: 6 sensor channels (GyroX, GyroY, GyroZ, AccX, AccY, AccZ), timestamps
• Data Quality: No missing values, no duplicates
• Sampling Rate: ~52 Hz
• Class Distribution: .....

Sensor Insights & EDA

Initial exploratory data analysis (Yuksel, 2021) reveals:

• High Variance & Noise: GyroZ exhibits the widest range (std = 12.0) and is the noisiest channel (51 identified outliers).
• Correlations:
  • Strong Positive: GyroZ ↔ AccY (r = 0.82). These channels are heavily coupled, likely representing yaw-related motion.
  • Moderate Negative: GyroX ↔ GyroZ (r = -0.46) and GyroX ↔ AccY (r = -0.42).
• Low Signal Value: AccZ hovers around -1.0 (gravity component) with low variance — limited discriminative signal.
• Outliers: 89 rows (8%) have at least one channel with |z| > 3. Handled via z-score normalization during preprocessing.

---

Training Results

---

Impact & Applications

Edge-deployed driving behavior classification

Tech Stack

Component	Tool
Model training	TensorFlow 2.21 / Keras
Data splitting & metrics	scikit-learn 1.8
Data processing	pandas 3.0, NumPy 2.4
TFLite inference	LiteRT
Visualization	Matplotlib 3.10
Target hardware	ESP32-S3 (Xtensa LX7, 512 KB SRAM)

Getting Started

# Clone and set up
git clone https://github.com/tylerrleee/driving-classifier
cd driving-classifier
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# Run the full training pipeline
python /driving_cnn_pipeline.py

# Outputs saved to /output/:
#   driving_cnn.keras, driving_cnn_int8.tflite, driving_cnn_model.h

---

References

• Amidi Afshine, Amidi Shervine. Convolutional Neural Networks. https://stanford.edu/~shervine/teaching/cs-230/cheatsheet-convolutional-neural-networks/

• Goh, Vik Tor; Jamal Mohd Lokman, Eilham Hakimie; Yap, Timothy Tzen Vun; Ng, Hu, 2021, "Driving Events", https://doi.org/10.7910/DVN/F5JZHF, Harvard Dataverse, V1

• Jamal Mohd Lokman EH, Goh VT, Yap TTV and Ng H. Driving event recognition using machine learning and smartphones [version 2; peer review: 2 approved]. F1000Research 2022, 11:57 (https://doi.org/10.12688/f1000research.73134.2)

• Yuksel, Asim Sinan; Atmaca, Şerafettin (2021), "Driving Behavior Dataset", Mendeley Data, V3, doi: 10.17632/jj3tw8kj6h.3