Driver Out-of-Position Detection Using Deep Learning Techniques

Raviteja ANUMULA NAGA

Master's thesis

Abstract:

In the automotive industry, ensuring driver safety is paramount when developing new technologies. Modern vehicles, equipped with advanced driver-assistant systems, can operate with little or no human intervention, which frequently places the driver in an out-of-position state. This situation increases risks for both drivers and passengers, as current safety features are designed based on the assumption that occupants are sitting upright, facing forward, with their hands on the wheel. This thesis investigates the importance of out-of-position detection and hands-on-wheel monitoring in the context of autonomous functions. The main focus of this thesis is to perform analysis on videos collected using Field Operational Tests (FOT) at Bayerische Motoren Werke (BMW) to detect driver out-of-position scenarios. The aim is achieved by developing a binary classifier to distinguish between in-position and out-ofposition states. Additionally, to enhance the analysis, a hands-on-wheel classification model with three classes no hands, one hand, and two handsis implemented using transfer learning approach. These models are constructed using annotated data obtained through computer vision techniques which are validated using statistical measures such as correlation. State-of-the- art validation and evaluation methods with appropriate metrics are employed to assess the trained models. Finally, the study establishes a robust and reliable framework for improved FOT analysis and enhanced driver safety in modern vehicles.

Abstract:

In the automotive industry, ensuring driver safety is paramount when developing new technologies. Modern vehicles, equipped with advanced driver-assistant systems, can operate with little or no human intervention, which frequently places the driver in an out-of-position state. This situation increases risks for both drivers and passengers, as current safety features are designed based on the assumption that occupants are sitting upright, facing forward, with their hands on the wheel. This thesis investigates the importance of out-of-position detection and hands-on-wheel monitoring in the context of autonomous functions. The main focus of this thesis is to perform analysis on videos collected using Field Operational Tests (FOT) at Bayerische Motoren Werke (BMW) to detect driver out-of-position scenarios. The aim is achieved by developing a binary classifier to distinguish between in-position and out-ofposition states. Additionally, to enhance the analysis, a hands-on-wheel classification model with three classes no hands, one hand, and two handsis implemented using transfer learning approach. These models are constructed using annotated data obtained through computer vision techniques which are validated using statistical measures such as correlation. State-of-the- art validation and evaluation methods with appropriate metrics are employed to assess the trained models. Finally, the study establishes a robust and reliable framework for improved FOT analysis and enhanced driver safety in modern vehicles.

Keywords

Driver Out-of-Position Detection, Deep Learning, Autonomous Vehicles, Advanced Driver-Assistance Systems (ADAS), Safety in Autonomous Vehicles, Computer Vision, Hands-on-Wheel Detection, 2D Human Pose Estimation, Transfer Learning, Binary Classification, YOLO (You Only Look Once), Field Operational Tests (FOT)

Language used: English

Date on which the thesis was submitted / produced: 20. 8. 2024

Thesis defence

• Supervisor: Dr. Christina Bauer

UNIVERSITY OF SOUTH BOHEMIA IN ČESKÉ BUDĚJOVICE

Faculty of Science

Master programme / field:
Artificial Intelligence and Data Science / Artificial Intelligence and Data Science