Overview / Usage

The Overall goal of this project is to create an android application which can turn an android device into a security system using the device's camera which can function without cloud connectivity and use Deep Learning to detect anomalies locally on the device. This would create an affordable security solution with less threat to data theft and cloud dependency, apart from this it would also make unused android devices useful as it can be setup as a security system easily and hence reducing the number of devices thrown in thrash and causing pollution just because they're not considered useful or outdated. This will also be helpful for people who can not afford CCTV cameras or high end security systems and to those people who live in places having low or no Internet connectivity like people living in village areas.
The main challenge is to create optimized algorithms which can run on android devices which usually don't have much computational power. Another challenge will be dataset creation and usage.

Methodology / Approach

The approach for creating such an application would be using live data from camera of the device and using it to detect anomalies. The model used would be inspired from MobileNet, a model which is well suited for such tasks on low computational power devices. The dataset used for training the model would be created using CCTV footages, this pre-trained model will be present with the initial application installation, further it would be fine tuned using face images present on the device such that if the model recognizes any face whose photo is not present on the device, it would consider it as high chance of threat and would become more attentive. When the model predicts a high chance of anomaly, it would alert some selected contacts through message and ring an alarm to alert nearby people. The device running this application would require constant power supply for better performance, it can work using it's battery in absence of power supply.
This application will provide an affordable security system which would not require Internet connection and cloud integration (unless video storage is required) and runs the application locally and hence reducing the chance of data theft and failure in case of bad Internet connectivity. This will also be helpful for places like village areas which have low or no Internet connectivity.
The pre-model will be implemented using Intel optimized tensorflow and Intel optimized Python and will be trained on Intel AI DevCloud. The application will be build in Android Studio which would also use Intel optimizations. The final deployment and fine tuning of model will be done by using Tensorflow as it is compatible with android devices.

Technologies Used

Intel AI DevCloud, Intel Optimization for Android (Celadon), Intel Optimized Tensorflow and Python

Repository

https://github.com/alishdipani/Digit-Recognizer

Other links

• MobileNet paper
• Summary of MobileNet