A novel framework for detection of motion and appearance-based Anomaly using ensemble learning and LSTMs

The variable time-dependent densities in crowd motion and several occlusions in real scenarios make the task of spotting anomalies very laborious. Also, an anomalous entity can be perceived to be non-anomalous from a different angle of perception. The proposed methodology introduces a novel framework for crowd anomaly detection at a patch level by integrating a thread of bi-directional LSTM for motion-based Anomaly and a thread of ensemble learning which uses pre-trained CONV-nets to learn appearance-based anomalies. A novel post-processing step to address the underlying false predictions using a threshold comparator system is also introduced. The proposed methodology is tested on the benchmark datasets, namely UCSD PED-1, PED-2, and UMN dataset, and it performs comparably with the other existing methods in the literature.