Surveillance-oriented trajectory-based anomaly detection: supervised vs unsupervised learning approaches

Transforming human understanding in visual data to electronic vision systems has been one of the aims in the field of artificial intelligence. The advancements in the digital cameras and computational capabilities have triggered the global demand in automated video understanding and anomaly detection especially in surveillance videos. Nevertheless, activities recognition and anomaly detection especially in public scenes remained a huge challenge for researchers due to countless factors such as poor quality footage, illumination, shadows, occlusion, noise etc. The state-of-the-art on behaviour analysis has built a bunch of sophisticated techniques for trajectory analysis. This thesis presents two trajectory learning methods; supervised and unsupervised trajectory learning and analysis in object's motion pattern. Investigation on object's motion trajectory for event recognition and anomaly detection frameworks is carried out for a video surveillance camera in an outdoor car park environment. The main challenge for outdoor environment videos is the variation in illumination. For a better outdoor object tracking process, an adaptive GMM background modelling is implemented for object trajectory extraction. Subsequently, for a better understanding of mobility data, a hybrid spatiotemporal trajectory modelling is implemented to abstract object's mobility data into different level; spatial location, velocity and acceleration. However, different type of object will have different repetition of motion pattern. Therefore, type of the object is classified before the motion pattern clustering. A scene modeling is done before the event learning system to analyse the point of interest of the motion trajectory. For supervised activity learning method, the scene is decomposed into regions and the trajectory pattern for each activity is grouped and clustered. An offline trajectory learning method is done by using Principal Component Analysis (PCA). The experiment result shows the accurate event recognition and anomaly detection. However, substantial amount of labelled training dataset has to be collected and event learning cannot be performed adaptively. Therefore, the work is extended to a second approach, unsupervised trajectory learning method to solve the addressing problem. The scene modeling is performed by clustering the entry and exit points of the object trajectories by GMM. In order to achieve automation in clustering trajectory pattern, a K-mean clustering algorithm is used to cluster the entire trajectories training dataset. A HMM framework is implemented in the system to devise parametric activity models for each class of activity. This system can perform the activities learning adaptively and fully automation. The comparison for both approaches is demonstrated in detail in this thesis.