In-Vehicle Driver State Analysis Using Image Segmentation

Abstract

This thesis investigates in-vehicle driver state analysis using deep learning, with the aim of improving road safety. The central idea is a data-driven approach that leverages semantic segmentation and monocular depth estimation to enrich existing datasets. First, a semantic segmentation model is trained to isolate the driver within the vehicle interior. Subsequently, experiments are conducted using autoencoders with various types of input image representations, including segmentation masks, estimated depth maps, depth sensor data, RGB images, and combined RGBD images. Three different autoencoder architectures are compared, each incorporating temporal information through the analysis of image sequences. Experimental results demonstrate that when driver masking is applied, anomaly detection performs better with estimated depth maps than with depth sensor data. These findings suggest the potential of semantic segmentation and monocular depth estimation to enhance anomaly detection performance in in-vehicle driver monitoring systems.