1. Moving Target Detection Using Infrared Sensors
Arisa Kitagishi, Babak Ebrahimi, Dr. Abhijit Mahalanobis
University of Central Florida
Introduction
Methods
Results
To provide better object detection using infrared sensors as not many computer vision research focus on frames obtained from infrared sensors. We would like to provide fast and accurate predictions using our network to be able to utilize it in real-life situations. 
Our goal is to achieve 80% or more in recall with less than 0.1 false alarm rate at long ranges (more than 4km).
We initially had background subtraction alone to see the results
FoveaNet
Then, we applied CNN to it to improve the results
               
               
Faster RCNN performed about 60% lower recall at false alarm rate of 1.1 compared to our method. Faster RCNN tends to have high false alarm rate when targets were detected. 
Without CNN
Without CNN
Background Subtraction
Results/ Experiments
Dataset
Data                         Ground Truth
                               Prediction
We plotted the
results using
constant false 
alarm rate to
eliminate
outlier caused by
the low threshold 
NVESDA ATR Development Dataset:
Taken using infrared sensors
Contains 1, 1.5, 2, 2.5, 3, 3.5, 4, 4.5, and 5km ranges
Contains night and day time folders
Each video contains 1800 frames
What we utilized:
2 furthest ranges (4.5km and 5km)
Night and Day folders
34 videos
The recall increased more than 15% at false alarm rate of 0.5 and even more for smaller false alarm rates by incorporating the CNN in the process. 
With CNN
Faster RCNN
CNN
References
Background Subtraction
1. Rodney LaLonde, Dong Zhang, and Mubarak Shah, ClusterNet: Detecting Small Objects in Large Scenes by Exploiting Spatio-Temporal Information, in Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2018), Salt Lake City, UT, June 18-22, 2018.