Automated Crowd Anomaly Detection and Localization using Video Analysis

Abstract

Anomaly detection and localization is a challenging problem in the field of computer newlinevision. Its objective is to infer the salient activities happening in a crowded newlineplace. This involves finding representations for object behaviors present in the newlinescene. Motion is an important attribute of video. The detail about the behaviors newlineis obtained from the collection of informative and meaningful appearance and newlinemotion features. These features may be captured from optical flow, spatiotemporal newlineand trajectory techniques. The present work is based on the understanding of newlineobject behaviors to distinguish between normal and abnormal activities. newlineOptical flow is widely used practice to detect the motion of objects. Magnitude newlineand direction of optical flow are obtained from the velocity of each pixel in the newlineframe. Histogram of magnitude (HoM), statistical and positional features of an object newlineare helpful to learn temporal and spatial characteristics. Statistical features are newlineused to differentiate among objects in the frame. Positional features are selected newlineto detect as well as locate anomalies from frames. Many times, it is observed that newlinethe foreground occupancy of object dominates over motion feature. An object with newlinemore area but less speed is anomalous. To include such behavior, influenced by newlinePhysics, a momentum feature is proposed to identify anomalies. Recently, deep newlinelearning features are also utilized to capture spatial and temporal features and employed newlineto detect anomalies. Convolutional neural network extracts deep features newlinewith its layered architecture. VGG16 pre-trained model is used to learn spatial appearance newlinefeatures of normal and anomalous objects. Two approaches are explored newlineto detect anomalies, homogeneous and hybrid approach. It is seen that a combination newlineof hand-crafted and deep features is more suitable to detect and locate anomalies. newlineDeep feature representation is attained over the raw magnitude of optical flow newlineusing stacked autoencoder. Autoencoder extracts high-level structural information newlinefrom motion magnitude to