Counting in Dense Crowds using Deep Learning Logan Lebanoff Mentor: Haroon Idrees Counting in Dense Crowds using Deep Learning
Ground truth=1567, Proposed=1590 Previous Work Ground truth=1567, Proposed=1590
Applications Concerts Political Speeches Rallies Marathons Stadiums Crowd management Safety/surveillance
Fully-Connected Layer Our Approach Deep Learning Convolutional Neural Networks (CNNs) Image Convolutional Layers Fully-Connected Layer Output Layer
Framework MatConvNet Matlab Caffe Python / C++ GPU
Pretrained models Instead of training the network from scratch, we fine-tune an existing model AlexNet 8 weight layers ImageNet-vgg-verydeep-19 19 weight layers
Training 50 crowd images Split into image patches Resize patches to 227x227x3 pixels Input to network Various convolutional layers, max pooling layers, and fully connected layers
Loss function Experimented on different loss functions for the last layer Classification problem Regression problem
Classification Network output a vector Softmax loss Size of vector is the max number of people in a patch For each index i, the value of the vector represents the confidence score that there are i people present in the image patch Softmax loss
Regression Network output a single number Experiments Represent the expected number of people in the image patch Experiments Euclidean Loss Sum of squares of differences Other
Localization Output the specific points where there is a face
Results
Results
Results
Results
Results
Results