OVERVIEW

The purpose of this project is to investigate and determine the efficiency of image classification deep learning models for both thermal and dash cameras. To maintain consistency between the two models, each trained model uses near identical scenes. This consistency makes comparative results more effective. 

Hypothesis

Predictions and Assumptions related to the investigation

Several interesting assumptions can be made related to the projects comparative results.

• Thermal imagery will be much more effective in diverse weather conditions.

  • Rain & Fog​

  • Snow

  • Nighttime

• Dash camera will be more effective in warmer environments and leverage its higher resolutions and FOV.

  • Warmer environments will be much closer to body heat, thermal images less pronounced.

• By fusing the models, with certain modifications to hyperparameters and layers within the overall model will optimize pedestrian and cyclist detection for automotive detection applications.

This project will create a unique dataset for varying weather conditions of pedestrians and cyclists in Ottawa. This dataset can be used by other projects, once published on a site such as Rotoflow.

By merging the training from both the thermal and pedestrian models, we will have many edge cases for general pedestrian and cyclist detection. The Indvidual weight modifications of the training parameters will optimize the final detection model for the purpose of a more uniform system.