A Discriminative Model for Multiple People Detection

Group activity recognition is becoming more important day by day for video surveillance, sports analytics, etc. Monitoring various cameras manually through human resources is a complex job; due to this, computer vision algorithms are being developed to perform lower and higher level tasks. This paper presents multiple people detection using the histograms of oriented gradients (HOG) feature descriptor algorithm through a support vector machine (SVM) based on the different group action class of persons. Multiple people detection for group action identification is a complex problem as accurate detection of individual persons requires extensive computation. To achieve multiple people detection for group activity, HOG feature extraction is proposed. HOG is precise and accurate person detection algorithm in the recent computer vision application. In addition, thresholding algorithm is implemented to collect the HOG feature vectors of definitely detected windows and firms the pathway followed by a person in the video frames. The proposed algorithm is evaluated through different aspects like group action categories and existence of occlusion over Haar and HOG features.