Dilated CNN Based Human Verifier for Intrusion Detection.

This paper proposes an intrusion detection algorithm for intelligent surveillance systems. The algorithm detects an intrusion threat via a dual-stage computer vision algorithm. In the first stage, the input of video sequences passes through a probabilistic change detector based on Gaussian Mixture Models to segment intruders from the background. The extracted foreground region is then passed through the second stage to verify if it is human. The second stage is based on a shallow convolutional neural network (CNN) employing dilated convolution. The system sends an alert if there is intrusion detected. The algorithm is validated and compared with a top-ranked change detection algorithms. It outperformed the compared algorithm on the i-LIDS dataset of sterile zone monitoring.