High-recall human abnormal behavior detection using MPEG-7 descriptor

Automated detection and recognition of human abnormal behavior is the key problem of monitoring systems. We construct a complete system that is able to alert the human operator when a knife in hand is visible in camera views. We use RealSense 3D camera to track hands, modified MPEG-7 EHD as feature vector and none-linear SVM as classifier. In this paper, we improve the feature extraction algorithm further by rotating feature vector and adding MGES feature methods, which are validated on the public knife detection database. Our hand knife detection algorithm achieves a considerable recognition accuracy of 92% and recall rate of 94.7%, increasing by 1% and 17% respectively, compared with similar studies. This improvement is pretty significant in strict security applications, which requires high recall and low false negative rate.