Open Collections

Deep learning approaches for crowd counting in images

University of British Columbia

Crowd analysis has been widely used in everyday life. Among different crowd analysis tasks, crowd counting is the most basic but essential one that measures the number of people in a particular area. Such counting and crowd density information is crucial to determine the maximum occupancy of a room or public area to address safety concerns. Counting by hand can yield an accurate number, but it is a tedious job and may not fulfill the time requirement for analysis. Therefore, automated crowd counting for accurately estimating the number of people in crowded scenes is needed. Deep learning has shown beyond human-level accuracy in many computer vision tasks. Recently, researchers have shown that deep learning methods can achieve the state-of-the-art performance in the task of crowd counting in images. The majority of these methods are based on density regression, where a density map is predicted and its integral is further calculated to obtain the final count. However, this learned density map is uninterpretable and could deviate largely from the actual person distribution even if the final count is accurate. Besides, existing crowd counting methods are mainly based on cumbersome feature extraction networks and can not be deployed in edge devices with limited computational power. In this thesis work, we proposed two models that tackle the above interpretability and computational efficiency concerns respectively: the joint crowd counting and localization model extends traditional counting only methods and provides precise localization results without additional model complexity; the ShuffleCount is a computationally efficient and accurate model that is trained through the specially designed knowledge distillation process and can satisfy the real-time crowd analysis requirement on edge devices. We evaluated both models by training and testing on public crowd counting benchmark datasets. Both quantitative and qualitative results are obtained and compared with existing state-of-the-art methods. Our superior results show the potential of deploying the proposed methods to real-life applications for efficient crowd analysis. The methods proposed in both JCCL and ShuffleCount can be generalized to other ideas to improve the interpretability and computational efficiency of the models.

Text

2021-04-22

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/77900

Electrical and Computer Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/