Visual detection and crowd density modeling of pedestrians

• 主要作者: Tan, Sing Kuang
• 其他作者: Cham Tat Jen
• 格式: Theses and Dissertations
• 語言: English
• 出版: 2017
• 主題: DRNTU::Engineering::Computer science and engineering
• 在線閱讀: http://hdl.handle.net/10356/72746
• 機構: Nanyang Technological University

This thesis attempts to address two problems that are related to the sensing and prediction of pedestrian distributions in urban settings. The first research topic is on the automatic collection ofpedestrian data, toaugment theinformationavailableto urbanplanners. Thesecond research topic is on automatically predicting the pedestrian density distributions given planned floor layouts of malls, potentially allowing architects to interactively adapt their designs and avoid excessively congested or underutilized regions. In the first part of the thesis, we will address on the problem of detecting pedestrians in camera images. The challenges faced now are large variations of appearances and poses, differences in illumination, occlusions and cluttered background. We tackle this by introducing a novel feature that captures second order intensity variations, which can complement existing HOG (Histogram of Oriented Gradients) and LBP (Local Binary Patterns) features. This has shown improvements in detection accuracy over some frequently used datasets. Visualization of ex- ample detection responses due to different features and weights are provided to more intuitively explain the reasons behind the improved performance. In the second part of the thesis, we model and predict the approximately steady-state pedes- trian density distributions in buildings. These are affected by a large number of latent variables such as the popularity of different shops and different possible routes that shoppers may take between shops. We proposed a probabilistic model that establishes the Markovian relationship between the different latent variables and parameters. We validated the predictions against ground truth pedestrian counts, and also analyzed how the predicted popularity of shops com- pared against measured traffic at shop entrances.