|dc.description.abstract||The Deep Learning Predicting Accidents project was completed during the Spring 2020 semester as part of the Computer Science capstone course CS 4624: Multimedia, Hypertext, and Information Access. The goal of the project was to create a deep learning model of highway traffic dynamics that lead to car crashes, and make predictions as to whether a car crash has occurred given a particular traffic scenario. The intended use of this project is to improve the management and response times of Emergency Medical Technicians so as to maximize the survivability of highway car crashes.
Predicting the occurrence of a highway car accident any significant length of time into the future is obviously not feasible, since the vast majority of crashes ultimately occur due to unpredictable human negligence and/or error. Therefore, we focused on
identifying patterns in traffic speed, traffic flow, and weather that are conducive to the occurrence of car crashes, and using anomalies in these patterns to detect the occurrence of an accident.
This project’s model relies on: traffic speed, which is the average speed of highway traffic at a certain location and time; traffic flow, which is a measure of total traffic volume at a certain location and time that takes into account speed and number of cars; and the weather at all of these locations and times. We train and evaluate using traffic incident data, which contains information about car crashes on all California interstate highways. This data is obtained from government sources.
The relevant data for this project is stored in a SQLite database, and both the code for data organization and preprocessing, as well as the deep learning model, are written in Python. The source code for the project is available at https://github.com/Elias222/DeepLearningPredictingAccidents.||en