Efficient Deep Learning Methods for Vehicle Incident Prediction

Abstract

The purpose of this thesis is to detail a step-by-step process of leveraging deep learning to solve an issue in the autonomous driving industry. We aim to predict the occurrence of traffic-related incidents using solely the visual information available in dashcam videos and deep learning models that prioritize efficiency. We propose a bottom-up approach, starting with the simplest models and building up from there. We begin with the most common deep-learning visual recognition system, the Convolutional Neural Network (CNN), and assess its ability to predict and identify incidents from single images. This method of single-frame prediction is surprisingly able to capture relationships between frames despite assuming independence between frames. As a logical next step, we add temporal structure to the model to capture the motion dynamics of our video data. We consider two approaches: 1) Linear Layer Feature Concatenation (LLFC) and 2) Recurrent Neural Network (RNN), and compare their performances side-by-side. We investigate the effect of changing sequence length, frame rate, and type of feature vector on the accuracy and efficiency of these two models. We find that the two models have comparable accuracies, but different evaluation times and feature vector preferences. Lastly, we explore our solution’s ability to generalize to a vaster scale of data and real-life settings.