Crowd estimation in images

• المؤلف الرئيسي: See, Chi Chao
• مؤلفون آخرون: Cham Tat Jen
• التنسيق: Final Year Project
• اللغة: English
• منشور في: 2014
• الموضوعات: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• الوصول للمادة أونلاين: http://hdl.handle.net/10356/58955
• المؤسسة: Nanyang Technological University

The idea of estimating sizes of large distant crowds in images taken from high mounted cameras is often seen as a difficult task in the field of computer vision. This project aims to investigate how various texture analysis methods and computer vision image processing techniques could be done to solve the problem. Through research, a multi output regression model for crowd counting was implemented for the sole purpose to reduce the over complications found in global and local regression models. The global model consists of a single model for counting, but the local model contains multiple regression functions and count localised spatial regions but becomes hard to scale if the model expands. However, the proposed regression model combines both approaches that involve a single model which can train multiple images and produced multi structural outputs. The implementation involves features selection of several low level imagery features such as segmentation, edge and texture which are inter-independent and are important for density estimation. These features can inherently produce a heat map where how it performs in weighted contribution and information sharing based on whole images and cells in partitioned images. A ground truth phase was manually dot annotated rigorously and forms the basis of comparison between the actual and expected counts in the training and testing set. Several statistical evaluation metrics were measured to test the bias and variance of the estimator model. The evaluation metrics would indicate the effectiveness of estimating crowd sizes. In future work, dynamic crowd structure segmentation could be implemented to improve the accuracy of crowd counting models, and other methods such as support vector machines and Maximum Excess over SubArrays (MESA) function could be recommended to extend the idea of crowd counting.