An Investigation of Videos for Crowd Analysis

Abstract

Day-by-day exploring biological diversity caused an exponential increase in population in the last few decades that necessitated the utmost need for safety, security, and management of crowd mobility in urban planning. Manual observing crowd dynamics is a complex enough process, which often produces the wrong results. Understanding crowded sequences greatly interests computer researchers because of its significant concern for research application fields. AI scientists have observed remarkable progress for the past two decades in developing the cognitive comprehension abilities of crowds in video surveillance using computer vision techniques. Posture/light variation, occlusion, shadow effects, and perspective distortion are intrinsic visual challenges leading to crowd analysis of complex phenomena, as these constraints limit the performance of computer vision (CV) mechanisms. newlineThis thesis introduces some self-solution-based deep-learning approaches to simulate crowd behavior in video surveillance. Four research objectives have been examined to investigate crowd dynamics, which focus on density estimation-based crowd counting, the development of automated crowd video surveillance development, and a human facial recognition model with useful motion information. The significant contributions of the thesis are as follows. newlineThe thesis begins with handling the occlusion issue for pedestrian estimation in dense crowd frames and proposed a dynamic kernel convolution neural network-linear regression (DKCNN-LR) model. This model works in two phases; first, a DKCNN model employs convolutional layers in such a way that the kernel weight of each subsequent layer is half the weight of the previous layer. Among all, the first three heavy kernel load layers identify distant camera regions (low-level) features. Later, the lighter kernel weighting layers help identify near-camera region (high-level) features. Second, a linear regression model is employed to perform parametric regression between the actual count (ground truth) and th