Neural network-based event detection for surveillance applications

We investigate neural network-based event detection for surveillance tasks. After signal segmentation and feature extraction, we take a first segment-based decision between the three classes calm, activity, and alarm. Higher-level decisions are taken subsequently from a temporal combination of multiple segments, which allows for the detection of predefined complex intrusion scenarioes. We describe the architecture of our neural network-based event detector for a fence security system and give evaluation results from field tests. Based on the different types of sensors of the surveillance system, data fusion techniques are used for joint processing of multisensor information to reach optimal results in terms of detection and false alarm rates. In addition, several approaches for in-field training with a small amount of adaptation data are evaluated to improve the classification performance for untrained environments.