Smart video surveillance for slow and metered connections

Traditional surveillance is done by recording video footage to a storage device continuously. However this generates enormous amount of data and reduces the life of the storage by continuous writes. Newer devices allow user to connect to the camera on demand from the Internet. One should consider bandwidth and cloud costs. Upload speeds are usually capped by Internet Service Providers (ISP) and cloud storage is expensive. Majority of data recorded by an indoor camera does not contain any information and motionless. When there is a motion it is most likely to caused by the household members, not an intruder. In this thesis we demonstrate a hierarchical cluster of cheap, disposable Raspberry Pi boards to monitor one or more cameras intelligently. System recognizes household members and does not save the data or trigger any action. When a face is not recognized locally, the image is uploaded to the cloud to perform further checks. If an image cannot be recognized by the cloud, it is marked as unknown. Bandwidth is only used when necessary and computation is performed by multiple devices simultaneously. We showed that our architecture can recognize $93.56\%$ of faces locally and $98.93\%$ in overall pipeline. Introducing an edge server locally can reduce the bandwidth as much as $93.56\%$. We also showed Raspberry Pi run containerized applications with minor overhead and performance drawbacks. Computation time is comparable to AWS Rekognition. Introducing a layered architecture increased the recognition rate by $5.37\%$ without adding significant network and cloud costs. The source code of this work is also available on Github under Mozilla Public License 2.0. Docker deployments are also open sourced. Docker images are publicly available on Docker Hub.