Traffic Signals and Neural nets

Last year, I got a chance to work at the University of Michigan Transportation Research Institute on a research project focused on using connected vehicle data and pedestrian data to reduce pedestrian-vehicle collisions.

This was carried out by analyzing data that we received from our test vehicles and pedestrian data from our test participants. The objective of the study was to derive insights from analyzing the data and make a decision to alert pedestrians/vehicles of an impending collision or a high risk area.

This data analysis was carried out using Long Short Term Memory neural networks. These neural networks are very good at making predictions based on time series data. This allows us to Fred in information from our connected vehicles including GPS data, accelerometer data, etc. The same is done for the pedestrian data we collect. As a result of this data analysis we are able to make a prediction about the likely location of vehicles in our test as well as pedestrian location.

From this data we are able to look for the likelihood that paths of these two classes will converge at the same time and are able to send warnings to the connected vehicle and/or the pedestrians.

The hopes is that this technology can be deployed at scale to ensure that autonomous vehicles in the future are able to take into account pedestrian data when making decisions about their movements and routes, minimizing the risk of collisions and damage to human life and property.