Real-Time Aerial Suspicious Analysis (ASANA) System for the Identification and Re-Identification of Suspicious Individuals using the Bayesian ScatterNet Hybrid (BSH) Network

Video monitoring and safety systems have been used to keep track of hostiles, conduct border control operations as well as to monitor the suspicious entities in public spaces. However, these systems are inadequate for the monitoring of large crowds due to the limited field of view of cameras. This paper introduces the Aerial Suspicious Analysis (ASANA) System for the Identification and Re-Identification of suspicious Individuals in large public areas using the Bayesian ScatterNet Hybrid (BSH) Network. The BSH network first estimates the human pose in each frame. Next, a batch of frames is used by the Bayesian 3D ResNext to identify individuals with suspicious postures. The system can also re-identify the identified suspicious individuals as they tend to move after committing the suspicious event. The proposed architecture is advantageous as it can learn meaningful representations quickly using the ScatterNet with fewer labelled examples. This is of great importance as real-life annotated training samples are hard to collect, especially for these applications. The pose estimation, suspicious individual identification, and re-identification performance of the proposed framework is compared with the state-of-the-art techniques. The proposed dataset is also made public which may encourage other researchers who are interested in using the deep learning technique for aerial visual crowd monitoring.