| Due to its widespread use, people counting has gradually become a hotspot in the research of the computer vision. People counting based on video only needs one-off investment, and it has high accuracy, good stability and other advantages. Indoor people counting especially video conference scenario is a challenging task, because video conference has complex background and recurrent occlusions .This article mainly discusses the research on feature extraction, machine learning and people counting based on video conference. The aim of this article is to finish people counting on a polling video conferencing media streaming. In this paper, the main work is as follows:Firstly, this article has an analysis and summary on the present methods of people counting based on video and points out their advantages and disadvantages. Considering the features of video conference scene, HOG feature is selected to be the feature in our passage.In this way, the research route of this paper is determined.Then, on account of the HOG which has the disadvantages on high dimensionality and single size, this paper presents a method to extract HOG feature based on sliding window strategy. And PCA is used to reduce dimensionality. So the HOG feature can fit with the video that has different sizes. At the same time, the efficiency of HOG also get an improvement.At last, this paper introduces the improved HOG to the people counting based on video session and uses MI-SVM as trainer to get classifier which has a higher efficiency.Meanwhile, combining to the feature of conference hall, this paper uses temporal and spatial characteristics to mark the people detected in the video and judges whether the people have occlusions according to inter-frame information. Finally, the paper realizes people counting on a real-time video stream.Through a lot of experiments and tests, this paper proposed a method which can realize real-time people counting in polling conference halls .And the method has certain practical value. |
PDF Full Text Request