1. Report 2
Brandon Silva

2. Relevant Papers
Multi-Target Detection and Tracking from a Single Camera in Unmanned Aerial Vehicles (UAVs)
Deep Learning for Moving Object Detection and Tracking from a Single Camera in Unmanned Aerial Vehicles (UAVs)
ClusterNet: Detecting Small Objects in Large Scenesby Exploiting Spatio-Temporal Information

3. Multi-Target Detection and Tracking from a Single Camera in Unmanned Aerial Vehicles
Uses classic computer vision techniques
Steps:
Estimate Background Motion using Shi-Tomasi corner detector
Detect Moving Objects by applying Lucas-Kanade method
Classify targets and track using Kalman filter
F Score
Background Subtraction
0.552
Target Classification Only
0.777
Target Classification / Tracking
0.866

4. Deep Learning for Moving Object Detection and Tracking from a Single Camera in UAVs
Uses same methods as in last paper for detection and tracking
Adds a simple CNN network that takes patches of detected motion, and classifies it as moving object or false alarm
This reduces false alarms and increases accuracy
Detection Accuracy:
Only Motion Difference
Deep Learning with Appearance
Precision
0.630 ± 0.11
0.819 ± 0.09
Recall
0.766 ± 0.15
0.798 ± 0.10
F-Score
0.684 ± 0.10
0.806 ± 0.08

5. ClusterNet
Use ClusterNet network to find regions of motion, and FoveaNet network to generate heatmaps for detected motion in wide area images
Uses 2D convolutional layers to preserve temporal relationship between the frames
The 2 frames before and after the central frame give enough information to classify motion accurately
This can can be applied to the Multi-UAV dataset, as UAVs are very small, where they cannot be seen by humans

6. Video Stabilization
Tried a few different methods for filling in area cropped out from stabilization
Uses SURF feature detector
30 Frame smoothing window is used for calculation camera motion

11. Challenges
Stabilized frames with annotations are slightly out of sync with ground truths for the videos
Needs to be fixed
Calculating correct bounding boxes may also be why they are off

12. ClusterNet on WPAFAB Results

14. Next Week Fix stabilization with the annotations
Generate heatmaps from bounding boxes
Try to apply Foveanet network to dataset using heatmaps for labels
Look at SSD/RCNN/YOLO for bounding box approach
Use classic computer vision algorithms to detect motion, and pass that data alongside frames into the network to help increase accuracy