D-CAD

Abstract

In this digital era, the complexity-efficient anomaly concentration becomes a major concern for the surveillance systems. This paper addresses the issue with an efficient and rapid technique for anomaly concentration and detection in the crowded scenes. Our model is trained by a concept of the deep learning (i.e. convolutional neural networks) framework, which makes use of an advanced feature-learning technique. Initially, a deep neural network is used for prematurely segmenting many normal frames. Our network works on the small frame being the initial stage, before cautiously rescaling the remaining features of interest, and assessing those at the following step using a more complex and deeper CNN. Superficial layers of the deep network (constructed as Gaussian classifiers, acting as weak classifiers) detect simple normal frames, such as background frames, and more complex normal patches are detected at deeper layers. It is shown that the proposed original technique (combination of deep neural networks and Kalman filter) will churn out comparable to today's top-performing localization and detection techniques on standard touchstones, but outshine those in general with respect to required computation time.