Browsing by Subject "abnormal event detection"

Ubiquitous surveillance has become part of our lives to increase security and safety. Despite the wide application of surveillance systems, their efficiency is limited by human factors, such as boredom and fatigue; because most of the time, nothing unusual happens. In safety-critical applications, time is essential and it is vital to act fast to prevent costly incidents. This thesis proposes a two-stage abnormal crowd event detection framework based on k-means clustering in the first stage, and sparse representation based methods in the second stage, to alleviate the laborious task of video monitoring. We conduct a literature review of 18 studies, where we specifically focus on sparse representation based methods. Accordingly, we choose the spatio-temporal gradient feature due to its simplicity, efficiency, and effectiveness in motion representation. After extracting features only from normal events, k-means clustering is applied to separate different motion feature clusters. Then, clusters with smaller samples, which are deemed to contain mostly abnormal features, are removed according to a threshold. In the second stage, we learn a dictionary for each remaining cluster using the approximate K-SVD algorithm. In testing, the reconstruction error of a feature against a learned dictionary and its sparse representation is used to determine an abnormality. We conduct extensive experiments on a standard dataset to evaluate the detection performance of the method. Furthermore, the effect of hyper-parameters in our method is investigated. We also compare our method with different methods to examine its effectiveness. Results indicate that our abnormal event detection framework can successfully understand abnormal events in a scene while running in real-time at 161 frames per second. With a few exceptions, no significant advantage of the two-stage sparse representation approach over a single large dictionary was found. We speculate that these results may be influenced by a small sample size. Nevertheless, our approach, due to its unsupervised nature, can be adapted to different contexts without additional annotation effort and using only normal events from videos. Therefore it motivates us for further development.