Crowd management and public space design become more and more important with population growth and world-wideurbanization. Crowd count or density is an important feature among the crowd information applied in this area, forcrowds of different density should receive a different level of attention. Accurate customer counting has become a key component as a holistic performance measurement for supermarkets, shopping malls, chain stores, museums, sporting venues, airports and so on. People counting and density estimation is a research direction taking practical application as the guidance, which makes this direction develop along with market demand. The key of research has developed from density estimation to people counting, the regression method from linearity to nonlinearity. The new methods proposed in this thesis are both based on motion information in a video and then optical flow is adopted to capture the information. The main contributions of this thesis include: 1. People counting using combined feature. The features adopted in previous methods are all extracted at pixel-level or based on local area, which are severely affected by factors such as occlusion. According to Holism, the properties of a given system are not only explained by its component parts but also the system as a whole. For people counting, we can model a crowd to extract macroscopic feature. The pedestrians perform different behaviors in crowds of different degree of crowdedness: when the crowd density is low, pedestrians are free to walk about, while the crowd density is high, they are hardly able to lift each foot. We can model the crowd with respect to these changes in behavior and apply optical flow to capture this motion information. Here we call the features used in previous work microscopic feature. Afterwards, we concatenate microscopic and macroscopic features into a combined feature vector and map it to people count using neural network. 2. Improvements on Mosaic Image Difference based foreground segmentation using optical flow. Foreground segmentation stage is a critical component for crowd density estimation models based on foreground segmentation. In places with people gathering and waiting, statistical background models can not segment accurate foreground because they progressively update as time goes by, this progress integrates stationary crowd (foreground elements) into background model simultaneously. To settle this problem, the Mosaic Image Difference (MID) ...
修改评论