Deep learning algorithms can be used to analyse and understand real-time crowd scenes. Exploring congested crowd patterns can enhance the understanding of traffic control, urban planning, public safety, and ecological surveys. A prominent application for this is to prevent disasters such as death and injuries during highly crowded events and religious pilgrimages. Furthermore, my project aims to efficiently extract semantic information such as crowd density estimation or counting, which can be used to predict crowds’ behaviour. This will be accomplished by developing a novel deep-learning model that improves the efficacy and accuracy of existing approaches. The project also involves training a modern machine learning model based on visual data collected using Unmanned Aerial Vehicles (UAVs) such as drones. This project will help me gain an in-depth understanding of the deep learning model and create a base and opportunity for me to explore this discipline further to solve more complex problems in the future.

Why is this problem special in ML field?

When people talk about machine learning, they either talk about classification, self-driving cars or regression-based algorithms. However, classification through image can be done easily by flattening the end vector to the number of classes you have in the data set. Unlike typical classification, crowd counting is special because it requires the algorithm first to detect if the object is person or not and second to produce the count. A classification will return true for the person being in an image, but it won’t give the count. So machine learning practitioners around the world came up with ideas to produce the end result as a density map and then calculating the sum of the heat map. I know it doesn’t make sense yet.

1. Consider the image with crowd:

An image showing dense population.

1. The image above has been annotated by the professional and provided the location of each individual in matrix form.
2. Through the use of matrix form and some Gaussian Preprocessing we can obtain a density map as such:

1. The sum of this density map will correspond to the number of people in the given image.

Now let the image denote the input parameter for the model and the density map to be the target.

If we do this, we can build an end-to-end convolutional neural network that will learn how to predict a density map when a new image is entered.

There are also other ways of detecting the number of people, such as by bounding boxes. However, the most effective models use the above as their target, because, here the model doesn’t face problems with image. Some problems include shadow, weather, scene, occlusion and much more.

Content

• Raw Links: