Crowd Counting

PROBLEM STATEMENT:

To develop a method that can accurately estimate the crowd count for an individual image with arbitrary crowd density and arbitrary perspective.

PROPOSED SOLUTION:

We use a multi-column convolutional neural network (MCNN) , which has proposed multi-column deep neural networks for image classification. Our model contains three columns of convolutional neural networks whose filters have different sizes. Input of the MCNN is the image, and its output is a crowd density map whose integral gives the overall crowd count.

TECHNOLOGY USED

The code was implemented on Google colab where we have used python as our programming language, we used ShanghaiTech dataset for training the model. The framework used was Keras.

METHODOLOGY

Phase 1 : The learning phase consisted of learning the basics of ML, DL and CNN. For that , we completed the Andrew NG course named “deeplearning.ai”. Also, for CNN we followed some videos on YouTube (MIT).
Phase 2: This consisted of implementing our knowledge about the subject on some basic projects such as classification of digits using the MNIST dataset.
Phase 3: We started with the Crowd Counting Project using Google colab. We have used the ShanghaiTech dataset which has a variety of related images with different perspectives. We used keras as the framework for our project.

RESULTS

The value of mae is 47.31868438720702

FUTURE WORK

In future we are planning to expand this into a real-time analysis on an actual crowd, the program currently finished can give an approximate count of the number of people in a crowd based on a static image.

KEY LEARNINGS

Main things that we learnt during the course of this project include getting to learn a lot about machine learning, deep learning and the way how images are preprocessed before being used for training the model. Also we learnt many new algorithms and made a suitable model using CNN.

CONCLUSION

Crowd counting using has a wide range of applications which include crowd control, disaster management, violence detection , etc.

REFERENCES

Following are the papers that we used as reference for our project:

1. [MCNN] Single-Image Crowd Counting via Multi-Column Convolutional Neural Network (CVPR2016)
2. [CSR] CSRNet: Dilated Convolutional Neural Networks for Understanding the Highly Congested Scenes

TEAM

● Shubam Agrawal
● Anish Patil
● Sasidhar Swarangi
● Rutwik Mulay
● Deepak