Abnormal Behavior Detection In Crowded Scenes

Situational awareness is a basic function of the human visual system, which has attracteda lot of research attention in machine learning, computer vision and related researchcommunities. Video surveillance of public and private space is fast becoming the norm intoday’s modern world with high-tech CCTV (closed-circuit television) cameras and imagecapture devices increasingly becomingaffordable, ubiquitous and with greater video-datasaving capabilities. There is also an increasing global demand for video-based intelligentsituation awareness systems. This demand stems largely from the need for better security,safety monitoring, and rapid response to dangers particularly when events or behaviors thatcan be contextually labeled as ’abnormal’,’strange’ or ’unusual’ occur within the observedscene. Applications where such systems are usually deployed are in: public transport systems(airports, sub-ways, and train stations), healthcare facilities, business areas, shopping mallsand banks, sport arenas, residential areas and so on.One unique quality of the human visual system is that it can very easily recognize,interpret and make high-level inference on object behaviors in different settings. With verylittle prior information, we can infer when a situation is transitioning from normal behavior toabnormal or unusual behavior. However, in order to aid humans to more effectively andefficiently perform this task, there is the challengeto build intelligent computer visionrecognition systems that can be used to support theenvironmental awareness function ofhuman vision. Such intelligent systems can visually observe and automatically extractinformation from raw video data for describing behaviors of moving targets, and learn todistinguish what is semantically meaningful to the human observer as ’normal’ and ’abnormal’behaviors.This thesis addresses three important aspects of the machine learning process for humanabnormal behavior detection. First, we propose a novel robust behavior descriptor forencoding the intrinsic local and global behavior signatures in crowded scenes. This descriptorwas applied to detect three types of abnormal crowd behaviors usually found under panicsituations namely: sudden accelerated motion (RUSH), sudden dispersion in different directions (SCATTER), escape towards an exit direction (HERDING). The descriptor is basednot only on the commonly used optical flow features but also flow dynamics that encodeinformation on both the local and global interaction between moving persons on the scene.Second, the task of visual codebook development for quantizing the extracted featureswas addressed using a bio-inspired quantization technique. The result forms thebuilding-block for coding extracted information based on the ’bag-of-features’ paradigm. Thisimportant phase saves computational time and effort in the building of the visual codebook byremoving guess-work or extensive empirical determination of the appropriate codebook sizesand representative features, which is the challenge when only the popularly used k-meansalgorithm is applied. Our use of the ant clustering algorithm synergized with k-meansalgorithm helps to overcome the limitations of the k-means algorithm.By using Bayesian topic modeling to capture the intrinsic structure of atomic activity andinteractions in the video frames, we effectively tune model parameters to detect as quickly aspossible, the transition from normal to abnormal behavior. Such early detection of thetransition between normal phase and abnormal phase is of crucial importance in time-sensitiveapplications. It provides a vital window of opportunity for human operators to deployappropriate response to the changing situation just-in-time.Experimental results and analysis of the proposed framework on two publicly availablecrowd behavior datasets prove the effectiveness of our method compared to similar earliermethods for anomaly detection in crowds with a very promising detection accuracy rates. Thisapproach is very much suitable for detecting particular events and trends in human behaviors,especially in sparse or crowded scenes.