Event Title

Session 2F: Pedestrian Detection using Convolutional Neural Networks

Session Number

Session 2F: 2nd Presentation

Advisor(s)

Dr. Miles Wernic, Illinois Institute of Technology

Location

Room A113

Start Date

26-4-2018 10:35 AM

End Date

26-4-2018 11:20 AM

Abstract

The purpose of this investigation was to devise an efficient and accurate algorithm that is capable of detecting pedestrians’ positions within frames of live footage. After exploring various models for pedestrian detection and weighing their advantages, we settled on using a convolutional neural network (CNN) as the basis for our algorithm. Using pre-recorded footage from the Chicago Police Department as well as the California Institute of Technology’s pedestrian image dataset, we are currently training a CNN to recognize pedestrians within 640 by 480 pixel still images. Currently, the model can predict (with decent time efficiency on a Nvidia GeForce 940M processor) whether a pedestrian is present within the image, but its accuracy rates are concerning due to a lack of pedestrian data from the CPD and the awkward viewing angle of the Caltech footage. We are in the process of transitioning our model to its final stage, which will involve a regional proposal network in conjunction with the convolutional neural network. After training this final model with ample and more suitable data, our final results will be presented at IMSAloquium.

Share

COinS
 
Apr 26th, 10:35 AM Apr 26th, 11:20 AM

Session 2F: Pedestrian Detection using Convolutional Neural Networks

Room A113

The purpose of this investigation was to devise an efficient and accurate algorithm that is capable of detecting pedestrians’ positions within frames of live footage. After exploring various models for pedestrian detection and weighing their advantages, we settled on using a convolutional neural network (CNN) as the basis for our algorithm. Using pre-recorded footage from the Chicago Police Department as well as the California Institute of Technology’s pedestrian image dataset, we are currently training a CNN to recognize pedestrians within 640 by 480 pixel still images. Currently, the model can predict (with decent time efficiency on a Nvidia GeForce 940M processor) whether a pedestrian is present within the image, but its accuracy rates are concerning due to a lack of pedestrian data from the CPD and the awkward viewing angle of the Caltech footage. We are in the process of transitioning our model to its final stage, which will involve a regional proposal network in conjunction with the convolutional neural network. After training this final model with ample and more suitable data, our final results will be presented at IMSAloquium.