Deep-learning driven exploration of automotive exterior design attributes

Abstract

With the current trend in the automotive industry of electrification and autonomous driving, new technologies and methods in assisting automotive designers are being applied to the vehicle design process. Nevertheless, the current automotive design process focuses on implementing a more efficient way to develop ideation sketches to 3D models that could be realized with hyper-realistic renderings while reducing the cost of developing clay models of full-scale design models. Though the change brought efficiency for designers, it lacks consideration for encouraging the creativity or productivity of designers who contribute significantly to the design process. The study proposes the potential usage of deep learning implementation to support designers’ day-to-day design practices. The study defines exterior design elements that determine the nature of automotive design while also defining methods to generate design datasets necessary for deep learning through elements essential to design. In order to explore how to utilize deep learning from a designer’s perspective, six design elements of the previously defined design elements were redefined, a design dataset was constructed based on this, and a base model was developed to extract design elements through YOLO v5. The created baseline model delivered 95.3% accuracy and 98.4% sensitivity in detecting and extracting design elements. Six design elements were extracted using the baseline model to create a preliminary design database. The study further proposes implementing an application for designers by showing how to utilize the extracted design elements through three cases of design applications using deep learning. In addition, the study explores ways to utilize numerical factors such as design evaluation and emotional evaluation in the design database. Nevertheless, the survey on emotional evaluation of vehicle exterior design was conducted through 40 designers, and the prediction model for emotional evaluation of vehicle exterior design was produced based on the results of the sensitivity evaluation. Finally, possible extension of vehicle design attributes and potential approaches in exploring design features from vehicle design in line with deep learning utilization methods that can contribute to the overall automotive design process are discussed based on the limitations of this study.