Estimating the physics of an automobile collision from images

Author

Manek, Hardik

Call Number

LE3 .A278 2021

Date

2021

Supervisor

Silver, Danny L.

Degree Grantor

Acadia University

Degree Name

Master of Science

Degree Level

Masters

Discipline

Computer Science

Affiliation

Computer Science

Abstract

Estimating the type and severity of an injury suffered in a car crash shortly after an accident is a challenging task. However, such estimates are important to the insurance industry. Using the change in velocity of the vehicle involved in an accident (Delta-V) as well as the Location of Collision (LOC) and the Principal Direction of Force (PDOF), the severity of the crash can be predicted, and eventually, the typeof injury suffered by the occupants can be estimated. Currently, the methods for measuring Delta-V are telematics, mathematical simulations, and manual estimation from images. In our research, we introduce a deep learning and computer vision based approach to estimate Delta-V using images of the vehicle involved in the accident; hence, streamlining the insurance claim process. In this research, the models are developed using Convolutional Neural Networks (CNNs) to predict Delta-V and classify LOC and PDOF. Since the prediction of Delta-V from car crash images using machine learning is an unexplored field, we prove that deep learning can be used to solve this problem by developing a proof-of-concept. We create car-crash simulations using the Rigs of Rods simulator, which uses a soft-body physics engine to simulate the vehicle’s motion destruction and deformation. The images from the simulations are used to develop the baseline models to predict the crash velocity. The results provide strong evidence that the convolutional neural network models can be developed to predict Delta-V from the images. Based on these results, we develop models using real-world collision images to predict Delta-V and classify LOC and PDOF. The resulting models predict Delta-V with a mean absolute error of 3.83 km/hr for frontal collisions and 3.49 km/hr for rear-end collisions by applying fine-tuning and image-processing. We demonstrate a Multiple-Task Learning (MTL) model to predict Delta-V and classify the location of the collision. The MTL model has a mean absolute error of 4.19 km/hr for Delta-V and test accuracy of 92% for classification of LOC. The Single-Task Learning (STL) model to classify the PDOF has a test accuracy of 34.44%. The final MTL model has a mean absolute error of 6.56 km/hr for Delta-V prediction and a test accuracy of 32.22% for PDOF classification.

Rights

The author grants permission to the University Librarian at Acadia University to reproduce, loan or distribute copies of my thesis in microform, paper or electronic formats on a non-profit basis. The author retains the copyright of the thesis.