This project presents an anomaly detection and its design implementation in Jetson Nano. We employ temporal reasoning network along with self-attention to better learn temporal dependencies across the video frames. Our anomaly detection scheme is proposed for crime detection by using the input of surveillance camera (CCTV). Note that the red block indicates the time of anomaly while the green bounding box indicates the normal activities.
i) employing ImageNet pretrained model for feature descriptor; ii) adopting Hilbert-huang Transform (HHT) to aggregate the deep learned features; and iii) invoking temporal pooling to capture dynamic information in each sub-action.
i) utilizing super resolution to provide better visual information to facilitate the network training; ii) using knowledge distillation to transfer the information from HR videos into LR videos; iii) proposing another stream network, composed of RGB, Optical Flow, and Trajectory spatial modalities; and iv) invoking a bidirectional self-attention module to manifest various temporal dependencies.
This project presents virtual reality for hajj. For muslim, hajj is very meaningful, however, it is very expensive because we need to go to Macca. Virtual reality is one of the solutions to resolve the issue. Various places in the hajj will be modeled in 3D resemble like the actual conditions then allowing users to interact in it. It likes looking around Masjid al Haram which there is the Ka’bah and Mas'a (where sa'i) inside. In this project, we build a Hajj virtual reality applications by using the Oculus Rift. Oculus Rift itself is a pair of glasses with a viewpoint reach 110o with First Person Camera mode.