Human and Group Activity Recognition from Video Sequences

A good solution to human activity recognition enables the creation of a wide variety of useful applications such as applications in visual surveillance, vision-based Human-Computer-Interaction (HCI) and gesture recognition. In this thesis, a graph based approach to human activity recognition is proposed which models spatio-temporal features as contextual space-time graphs. In this method, spatio-temporal gradient cuboids were extracted at significant regions of activity, and feature graphs (gradient, space-time, local neighbours, immediate neighbours) are constructed using the similarity matrix. The Laplacian representation of the graph is utilised to reduce the computational complexity and to allow the use of traditional statistical classifiers. A second methodology is proposed to detect and localise abnormal activities in crowded scenes. This approach has two stages: training and identification. During the training stage, specific human activities are identified and characterised by employing modelling of medium-term movement flow through streaklines. Each streakline is formed by multiple optical flow vectors that represent and track locally the movement in the scene. A dictionary of activities is recorded for a given scene during the training stage. During the testing stage, the consistency of each observed activity with those from the dictionary is verified using the Kullback-Leibler (KL) divergence. The anomaly detection of the proposed methodology is compared to state of the art, producing state of the art results for localising anomalous activities. Finally, we propose an automatic group activity recognition approach by modelling the interdependencies of group activity features over time. We propose to model the group interdependences in both motion and location spaces. These spaces are extended to time-space and time-movement spaces and modelled using Kernel Density Estimation (KDE). The recognition performance of the proposed methodology shows an improvement in recognition performance over state of the art results on group activity datasets.