People detection in video streams using background subtraction and spatial-based scene modeling

Nowadays, the people detection in video streams is a topic of great interest in applications like video surveillance, customer behavior analysis, activity recognition, among others. Commonly, real world scenarios expose conditions where multiple people interact with each other, so, the people detection becomes more complex. Several approaches have been proposed to detect people from crowds, most of them focusing on the improvement of the characterization process. From them, the features based on Histogram of Oriented Gradients (HOG) have exposed superior performance. The majority of the people detection approaches perform a sweep along the whole scene looking for regions that can be classified as persons. However, this kind of heuristic searches tend to be time-consuming, even more if the characterization and classification have high computational costs. To cope with this, we propose to use background subtraction techniques to restrict the search of candidate regions to be classified as persons only over the foreground regions. Additionally, we include information about the scene spatial model in order to spread candidate regions in a more efficient way. The performance of our approach is assessed in terms of computational cost and accuracy by comparing against the people detector of the OpenCV library. To this, video records from real world scenarios drawn from public datasets are employed. As future work, proposed strategy will be tested on real world video surveillance systems for human activity analysis.