Vehicle sensors data analysis for detection of driver styles in braking maneuvers

I aimed to uncover and analyze braking patterns in driving behavior using machine learning and Controller Area Network-BUS (CAN-BUS) data. My primary goal was to understand how drivers employ the brake pedal across different scenarios, including highways, rural areas, and urban settings. I followed a structured approach, starting with data preprocessing and feature selection, leading to the development of a Variational AutoEncoder (VAE)-based deep learning framework. I carefully selected input signals like speed, steering angle, and accelerations for their consistency across various vehicle models. Streamlining the model's complexity and enhancing its interpretability were key considerations. During data preprocessing, I addressed challenges in resampling signals to a common 5 Hz frequency, applying noise-cleaning techniques such as interpolation and filtering. Segmenting driving behaviors by environment was crucial. This segmentation allowed me to explore how braking behavior differs across driving scenarios. The core of my project was the VAE, a generative model capable of extracting latent data representations. The encoder-decoder architecture learned and reconstructed driving patterns. To bolster the model's robustness, I included a classifier to distinguish real braking data from noise within the latent space. Training involved typical VAE loss functions, including reconstruction and KL divergence loss. The trained VAE transformed complex braking data into compact vectors. Clustering similar braking scenes was a challenge, but I optimized cluster numbers by minimizing variance in key performance indicators (KPIs). Insights from clusters were extracted using t-SNE visualizations, speed and acceleration plots, and KPI percentiles.