Photo: Siddharth Kankaria / Research Matters
Imagine you are in a crowded cricket stadium enjoying the game. There are thousands of enthusiastic fans around, just like you. What would happen if there were to be any untoward incident like a fire? Panic, chaos and worse, stampede, may ensue. Managing crowd and planning for disaster contingency at such heavily crowded events are always a challenge for civic authorities. In today’s world powered by social media, mobilizing a crowd in a short time for any event is no more a challenge.
Given that so many lives are at stake, how should the civic authorities plan for safety? How do we know if there is more crowd than the security and civic agencies can handle? Though these questions seem difficult to answer, technology can help here, say scientists from the Indian Institute of Science, Bangalore. In a recent study, Prof. Venkatesh Babu and his team from the Department of Computation and Data Sciences, have proposed a novel method that can accurately estimate the number of people in densely crowded gatherings. Powered by neural networks, this technique can use both images and video streams to count the crowd, claim the researchers.
Neural networks are computational models inspired by biology that enable a computer to learn from observational data, just like the human brain. They look at thousands of sample data and learn to recognise patterns. Hence they are used in applications involving image recognition, speech recognition and natural language processing. Convolutional Neural Networks (CNNs) are a subset of neural networks that are better suited to process and recognize images. “A typical CNN consists of multiple layers of neurons forming a hierarchy of pattern detectors like vertical and horizontal edges, colors, textures etc. This architecture is inspired from visual cortex of the brain”, says Prof. Babu.
The researchers have used CNN in their study to estimate crowd counts. “In our study, the model looks repeatedly at scenes with crowds and learns a ‘concept’ for detecting a person that is agnostic to the pose or scale of a person”, explains Prof. Babu. In other words, looking at many data with people, the model starts to identify people and differentiate them from other objects. “It also learns that people are more likely to be found against a grey background (road/pedestrian way), rather than a blue/green region (sky/water or foliage)”, he adds.
The most pronounced challenge in crowd counting is the variable crowd density – a crowd can swell in no time and parts of the same crowd can vary with the number of people. Other challenges include people merging with the background, varying camera position and perspective, and varying scale of the image. These different attributes call for different ways to measure the number of people, prompting the use of different CNNs for different types of crowd.
While there are many methods proposed to estimate crowd numbers, the advantage this method has over earlier methods is that it addresses scenes with varying crowd density very well. “Our technique uses an expert CNN to choose the best CNN from a group of CNNs. This helps it employ different CNNs for different parts of the scene. And our training method achieves this automatically”, remarks Prof. Babu.
The researchers tested their method with standard crowd-scene datasets with each containing a multiple images of dense and sparse crowd scenes. As reference they used the crowd count for these images calculated through annotations. They measured the mean absolute error to indicate accuracy and the mean square error to indicate robustness. The switch-CNN method proposed by the authors performed exceptionally better than earlier schemes for scenes with high density variations.
Funded by the Science and Engineering Research Board (SERB), Department of Science and Technology (DST), Government of India, Prof. Venkatesh Babu and his team are set to present this work in the "Computer Vision and Pattern Recognition" (CVPR) conference in July'17 at Honolulu, Hawaii, USA.
So the next time there is a controversy like that of the presidential inauguration of Donald Trump, perhaps the news reports should look at backing their numbers with evidence from such smarter algorithms.