South Africa has a wealth of diverse animal life which defines our country. We are known for big five which attracts many tourists from all over our world who want to see these walking natural wonders.
Unfortunately humans can sometimes present a threat to this natural habitat through the illegal hunting of endangered species. This practice is known as poaching and can potentially lead to the extinction of an entire species. This has been witnessed before in history with the case of the Dodo bird and the Tasmanian tiger.
The worst poaching epidemic in South Africa is the poaching of Rhinos. With a population of 20 000 Rhinos, South Africa has become the primal target of Rhino poaching. There has been an increasing trend in poaching, according the Kruger National park, about 30% more events have occurred since October 2015.
National parks attempt to preserve the natural habitat which these wild animals reside in, with a well-known example being the Kruger National park. In the parks, rangers are responsible for the safety of the animals. The rangers employ various technologies to enhance their awareness of the environment.
This project proposes a low powered threat detection system that also does not require a high amount of networking traffic. The system will constantly be monitoring a video camera overseeing an area of the park and attempt to identify possible threats and then send an alert to a ranger to take notice of the possible threat.
A possible threat will simply be classified as a person detected in the frame, since the cameras will be placed in locations where only wild animals are expected to roam. (Thus a person is an unusual occurrence which should be detected in this circumstance) Additionally more threat categories could refine the system even more if desired, such as adding categories for weapons.
This project is focused on practicality and showcasing a good use case of an IoT system which employs deep learning in an environment that introduces limitations. The limitations being power availability and networking. Power availability is addressed by designing the system to use minimal power, such as using the Movidius Neural Compute Stick for inference. Networking is addressed by having the deep learning performed on the device itself and only transmitting images when a video frame is classified to have importance by the deep neural network. Doing the inference on-site removes the dependency of using cloud infrastructure which requires a large amount of networking (since every frame would then have to be transmitted for processing) which cannot be relied on always be accessible in a remote area where signal strength to the intended station could be low. (Thus limiting networking bandwidth)
The whole system should be fully self-contained once deployed. Power is generated by solar panels which feed through a charge controller which takes care of charging the battery using the solar panel. The battery power is then converted to power a single board computer such as the Raspberry Pi or Intel Joule. Connected to the computer is a webcam for the video feed and a Movidius Neural Compute stick to perform low power inference on the video. The single board computer should have at least two forms of networking to alert when a threat event is detected. Two forms of networking are chosen to have redundancy as it might needed to potentially transmit critical information. (More reasons are given later)
The initial versions of the system will make use of the inception V3 model which will be retrained to new categories. Should better performance of detection be required, alternative architectures can be explored.
The system will depend on the Movidius Neural Compute Stick which will allow it to run low power inference on the image frame in order to find possible threats. This will allow the system to run off battery and solar power and not require a high data rate connection to constantly send image data. Only if a possible threat is detected is an image sent. These design choices allow it the system to be placed in a remote area and be fully self-contained without having to put heavy consideration into power and networking requirements.
Should a threat be detected, a signal can be transmitted via a low bandwidth radio transmission and a GSM modem. (Optionally also sending the picture of the suspected threat) Having multiple sources of transmission can allow for redundancy in the case of signal jamming. Additionally a heartbeat signal which periodically gets sent from the system can indicate that it is still online and not under threat of jamming. (If this signal stops after a set amount of time, a ranger will be alerted to investigate)