Drone Surveillance

Pedestrian detection from aerial view and neural network compression

Content

1. Neural Network Compression
2. Pedestrian Detection from aerial view

1. Neural Network Compression

• By making architectural changes eg. SqueezeNet
• By doing compression post training eg. Deep-compression

By Making architechtural changes

Observations

• Maxpool over average pool
• Decrease in number of feature maps deep in the architecture
• Number of feature maps per layers depends upon the type of dataset chosen
• Fire models should be implemented later in the network
• Replacing 5x5 with two 3x3 results in slight drop in accuracy

Experiments

By making architechtural changes in reference to SqueezeNet

Results

2. Pedestrian Detection from Aerial View

Dataset Description

We have used stanford drone dataset : http://cvgl.stanford.edu/projects/uav_data/.

• Sampled frames from different scenes in a deterministics approach
• Created subsets of the dataset consisting of ~100, ~1000 and ~2000 video frames
• ~100 image samples: consisting of ~2144 objects in total across all images
• ~1k image samples: consisting of ~15545 objects in total across all images
• ~ 2k image samples: consisting of ~19926 objects in total across all images
• Corresponding test sets also contain the respective number of images

Results

Loss convergence graphs

Demo Inference Videos

Scene 1

Scene 2