Detection of abnormality in video crowded scenes is highly challenging in computer vision over the past decades. We propose a method for the detection of abnormal events in surveillance video sequences. In this method, we combine the spatial-temporal convolution neural network (CNN) with handcrafted feature sets such as Histograms of Optical Flow (HOF) and Histogram of Oriented Gradients (HOG) for anomaly detection in contiguous video frames. Handcrafted features learned sparse by using our novel method which we named it Iterative Weighted nonNegative Matrix Factorization (IW-NMF) is based on sparse NMF. These feature extracted from active volume cells that included moving pixels to reduce computational costs. The architecture of the CNN model allows us to extract spatial-temporal features and using handcrafted features to increase the accuracy of detection and to ensure robustness against local noise. we evaluate our framework on popular datasets that contain abnormal activites. our method results are better than most of the other method. and achieves a very competitive detection performance compared to state-of-the-art methods.
Paper
Full text
Deep and Sparse features For Anomaly Detection and Localization in video
Semantic Scholar · Computer Science · 2019
Abstract
Detection of abnormality in video crowded scenes is highly challenging in computer vision over the past decades. We propose a method for the detection of abnormal events in surveillance video sequences. In this method, we combine the spatial-temporal convolution neural network (CNN) with handcrafted feature sets such as Histograms of Optical Flow (HOF) and Histogram of Oriented Gradients (HOG) for anomaly detection in contiguous video frames. Handcrafted features learned sparse by using our novel method which we named it Iterative Weighted nonNegative Matrix Factorization (IW-NMF) is based on sparse NMF. These feature extracted from active volume cells that included moving pixels to reduce computational costs. The architecture of the CNN model allows us to extract spatial-temporal features and using handcrafted features to increase the accuracy of detection and to ensure robustness against local noise. we evaluate our framework on popular datasets that contain abnormal activites. our method results are better than most of the other method. and achieves a very competitive detection performance compared to state-of-the-art methods.