Crowd Counting Based on Deep Learning

Crowd Counting Survey

This is a brief survey about crowd counting. Firstly, I will given some related papers with my comments from both deep learning and traditional perspectives.
Then I will list some public available datasets and benchmarks.

Finally, I will give the most feasible method to predict the crowd density and explain the reasons.

Paper List

• Deep Learning based Methods

  1. MULTI-SCALE CONVOLUTIONAL NEURAL NETWORKS FOR CROWD COUNTING

  2. Single-Image Crowd Counting via Multi-Column Convolutional Neural Network

  3. CrowdNet: A Deep Convolutional Network for Dense Crowd Counting

  4. Cross-Scene Crowd Counting via Deep Convolutional Neural Networks

Dataset

There are different ways to count the crowd. One is to predict the crowd count
directly via body detection or face detection. Using this way need to train dataset with face annotation or body annotation (some boxes). This way is hard to count for some images with large crowd. Another way is to generate a heat maps of the people. Then, the count of the crowd is the integral of the heat map. Predicting in this way need to train dataset with the head center point annotations(some points). For both the two types, there some public available
datasets, links to these dataset are listed follows:

• Datasets with head annotations

  1. UCF_CC_50

• Datasets with person annotations

  1. UCSD Pedestrian Dataset

Dataset Preprocessing

In this section, I mainly focus on how to process the dataset with head annotations. The process consists of two steps, one is to generate ground
truth head maps. Another is data augmentation.

• Generating Ground Truth Heat Maps

  Here I describe the method used in 4 in paper list. If there is a head at
  pixel \(x_i\) , we can denote this as \(\delta (x-x_i)\). Hence an image with \(N\) heads labeled can represented as a function:

  $$H(x) = \sum_{i=1}^{N}\delta(x-x_i)$$

  To convert this to a continuous density function, we can convolve this
  function with a Gaussian kernel \(G{\sigma}\) . So the density can be
  formulated as \(F(x)=H(x)*G{\sigma}(x)\) . If you want to know more
  details, please refer to my Project Page

• Data Augmentation

  For data augmentation, we follow the method used in 3 in paper list. Firstly, we generate different scales images. Then, we crop 225x225 patches with 50% overlap from those different scales images.

  Some visualization:

  

Network Building

To be Continued.