Anirudh Madhusudan, Harshad - PowerPoint Presentation

Slide1

Anirudh Madhusudan, Harshad Rai, Kriti Gupta, Sarvesh Rajkumar | Faculty Advisor: Prof Richard SowersDepartment of Industrial and Systems Engineering, College of Engineering, University of Illinois at Urbana-Champaign

Machine Learning for Safer Roads

NEW YORK CITY VISUALIZATION 

INTRODUCTION Vision Zero is a multi-national road traffic safety project that aims to achieve a system with no fatalities or serious injuries on roads. Through the Vision Zero initiative, New York City has been documenting information on all traffic collisions within the cities over the past few years. These datasets are intended to help the government authorities implement treatments to reduce traffic fatalities/injuries in the future.The NYC Traffic Collision Dataset provides interesting insights on motor vehicle collisions with over 1 million data points. Each data point bears several relevant attributes such as date, time, location of collision, cause of collision or types of vehicle involved. This poster will explore the frequency of collisions with regard to temporal variations, vehicle involvement and preceding events. This poster will also discuss a machine learning implementations that predicts the likelihood of a fatality, given a collision. Due to low event rate of traffic fatalities (less than 0: 1%) additional data-points are generated to help train the model better. This is achieved through the Synthetic Minority Oversampling Technique (SMOTE).

MACHINE LEARNING FOR SAFER ROADS

SMOTE

For the machine learning approach adopted, a very small amount of the collision data led to a fatality -( 1 fatality in every 10000 collisions). Due to this imbalance, the ML process is difficult and inaccurate.SOLUTION?Synthetic Minority Oversampling Technique (SMOTE) was adopted to handling this imbalance by creating synthetic samples of the fatalities. The algorithm selects two or more similar instances from the minor class (using a distance measure) and perturbs an instance one attribute at a time by a random amount within the difference to the neighboring instances.

CONCLUSIONSPrecision and Recall of 84% - 4 Layer Neural Net on NYC Dataset81% - Logistic Regression on NYC dataset

Upon filtering out high frequency collision spots that recorded more than 60 collisions, and then observing the geographical features of the spot on Google Earth, it was observed that 40% of them occurred near 4 way intersections and 80% of the High frequency accidents occurred on or near a ramp(bridge). This can b further subdivided into 2 categories namely:On or near the entry of a bridge(expressway).Just outside the underpass of a bridge.

Interesting trends observed in case of fatalities are, the number of fatalities tend to peak around 2:00 to 3:00 due to Alcohol involvement, Driver Inattention and Traffic Control Disregard. Driver Inattention and Traffic Control Disregard may be attributed to the early hours.

ML Implementation

SCOPE OF FUTURE WORK

Given the network topology around the collision spot, we plan to predict the likelihood of a severe injury or fatality. We then intend to identify the optimal network size that will reduce the error in this prediction.

The trends show that the collisions and fatalities are more likely during the night time.

The points on this map represent locations where the sum of number of persons killed or injured were greater than 5. From this it is noticeable that some spots in BRONX and of BROOKLYN have higher density of such accidents, which imply more focus is to be given to those areas.

Seasonality trends seem to be fairly consistent, with most collisions (and most injuries) being recorded at 17:00 to 19:00 which may be attributed to the rush to return home from work.

Injuries also seems to follow the trend of collisions, but having two distinct peaks(one at around 8:00 and one at 17:00 – 19:00), which can be attributed to people rushing to and from work.

University of Illinois at Urbana-Champaign

The effects of seasonality on collisions, injuries and the trends of fatalities for various causalities over time has also been studied and inferred upon. Ideas for treatments that reduce the number of collisions may be gleaned from the trends presented herein.