Essay On Crowd Analysis

2.2.1 Crowd Analysis Techniques Based on Computer Vision Computer vision methods have been widely used to study different aspects of crowd analysis. Jacques, Musse and Jung (2010) discussed three aspects: crowd counting/density estimation, tracking in crowded sights and high-level crowd behavior understanding.
2.2.1.1 Crowd counting Crowd counting is one of the fundamental issues in the area of crowd analysis as counting assists in more efficient crowd management such as overcrowding anticipation and managing future crowds. One of the famous counting approaches is counting-by-detection, which uses a trained detector using local features (e.g. histogram oriented gradients [HOG] or Haar wavelets) to locate individuals in the scene

A regression model is used to map holistic image patterns to the total number of people in the image. The process of counting-by-regression follows a standard pipeline. First, performing geometric correction to address the problem of perspective distortion wherein objects close to the camera looks larger than far ones. Then, extracting low-level features such as foreground segment (e.g. Gaussians-based background subtraction), edge (e.g. total edge pixels, edge orientation, and Minkowski dimension), texture (e.g. gray-level co-occurrence matrix [GLCM] and local binary pattern [LPB]), gradient (e.g. HOG and gradient orientation co-occurrence matrix [GOCM]) or a combination of multiple features. Finally, training a regression model such as kernel ridge regression (KRR) or support vector regression (SVR) to predict the total count based on the normalized features (Loy et al. 2013). However, Kurilkin and Ivanov (2016) argued that training several cameras is not feasible due to prospective variance and using learning algorithms for one camera affects the accuracy negatively. Moreover, lighting variations have a remarkable influence on the counting