Research on convolutional neural network for object classification in outdoor video surveillance system

Abstract

Nowadays indoor and outdoor video surveillance systems are very widespread. Earlier, in the days of the first surveillance systems, the processing power allowed only monitoring and recording the surveillance footage, but now it becomes possible to use various methods of video analysis; in this article, we investigate the convolutional neural networks application in the objects classification. In our previous work we developed the outdoor video surveillance system for detecting objects using fixed and PTZ cameras. System provides detection of moving objects with low computational cost and high accuracy. This paper summarizes the results of work on the existing outdoor video surveillance system for detecting objects and our new convolutional neural network object classifier based on the Keras and TensorFlow packages. Reliable determination of the object type allows the system to make decisions on processing the object information. The considered classifiers allow performing both simple classification (a person/not a person) and more complex one (error/person/car/animal) with insignificantly lower reliability. Object tracking in consecutive video frames can remarkably reduce the number of classification operations, because there is no need in performing them for each frame in case the object class has been identified with enough reliability. In addition, the integration of the developed networks into the existing video surveillance system is briefly described.