Long term carefully learning for person detection application to intelligent surveillance system

• Authors: Nguyen Dang Binh
• Venue: Proceedings of the 9th International Conference on Advances in Mobile Computing and Multimedia
• Year: 2011

In this paper we introduce a framework for unsupervised learning visual object detector in long sequences of continuous video data for intelligent visual surveillance system. The main idea is to (1) minimize the manual effort when learning a classifier and to combine the power of a discriminative classifier with the robustness of a generative model; (2) to exploit a huge amount of unlabeled video data by being long term and careful in selecting training examples; and (3) to start with very simple detection system using motion detection an initial set of positive examples is obtained by analyzing the geometry of the motion blobs. If a blob fulfills the restrictions the corresponding patch is selected. Negative examples are obtained from images where no motion was detected. Starting from these data sets a first discriminative classifier is trained using online boosting for feature selection [1] learning and applying a generative classifier using Principle Component Analysis (PCA) [2] to verify the obtained detections and to decide if a detected patch represents the object-of-interest or not. As we have a huge amount of data (video stream) we can be very long term and careful to use only patches for (positive or negative) updates if we are very confident about our decision. Applying these update rules an incrementally better classifier is obtained without any user interaction needed. We demonstrate the framework on a surveillance task where we learn a person detector by using this approach we avoid hand labeling training data and still achieve a state of the art detection rate.