Murdoch University Research Repository

Welcome to the Murdoch University Research Repository

The Murdoch University Research Repository is an open access digital collection of research
created by Murdoch University staff, researchers and postgraduate students.

Learn more

Novel architectures for spectator crowd image analysis

Jan, Yasir (2020) Novel architectures for spectator crowd image analysis. PhD thesis, Murdoch University.

PDF - Whole Thesis
Download (2MB) | Preview


Crowd image analysis can involve various tasks, such as head detection, head pose estimation, and body detection. These tasks face a range of issues including low resolution, varying crowd density, overlapping bodies, and image corruptions. Additionally, the techniques performing these tasks have computational overhead as well.

Existing head detection techniques perform poorly for very low-resolution images. They also cannot perform head detection and head pose estimation simultaneously for multiple heads. To address these issues, we propose a novel neural network architecture (WNet), which performs joint tasks on low-resolution multiple heads. Experiments on the spectator crowd dataset (S-HOCK) show that fewer images can be used for the simultaneous tasks.

Existing body detection techniques rely on the visibility of body parts. But in dense spectator crowd images, the bodies are occluded, and not visible clearly. We propose “pixel matching based body detection” (PMBD) technique, to reduce the effect of occlusion in body parts. It can locate the body region by color matching and proximity. Experiments are performed on the S-HOCK dataset to accurately detect occluded bodies in a crowd image.

When low-resolution images, such as spectator head images, get distorted with corrup-tions e.g. blurriness, pixelation, and fog, they are poorly classified by the neural networks. Existing robustness techniques aim towards a specific type of corruption and improve the accuracy. We propose a novel technique, called “Edge to edge scanline smoothing” (ESS), for image enhancement to mitigate the effect of a wide range of corruption.

Training of neural networks is a computationally extensive task. The proposed structure of neurons called the “Multiple Output Neurons” (MON), in contrast to conventional single output neurons, can be used in the training process to reduce computation in the neural networks. We demonstrate a reduction in computation for various tasks, such as spectator crowd head pose classification.

Item Type: Thesis (PhD)
Murdoch Affiliation(s): Information Technology, Mathematics and Statistics
Supervisor(s): Sohel, Ferdous, Shiratuddin, Fairuz and Wong, Kok Wai
Item Control Page Item Control Page


Downloads per month over past year