The beehive outside
The front of the beehive
The top of the beehive
Our neumorphic dashboard client
Simple reinforcement learning built into the experience. As people use the system, the system will improve seamlessly
Result of object detection model drawing bounding boxes on a test image.
Computer vision model sample size and accuracy.
Twilio SMS notification sent when sensor is triggered.
domain.com order confirmation
Those pesky murder hornets just can't bee-hive themselves, their just a bunch of wanna-bees! We were buzzy all day and night working on this project, and it's not too shab-bee. We aren't pollen your leg! Oh, and even though we used US Bees though out the project, the implementation should work anywhere in the world.
As if a global pandemic wasn’t enough to handle, murder hornets have arrived in the United States. It only takes a few of these guys to wipe out an entire colony of bees in just a few hours, decapitating the powerless honeybees and taking their thoraxes to feed their young. At a time when honeybees are already severely threatened, murder hornets jeopardize not only bee colonies but also our food system. In fact, one-third of the food we eat depends on pollination by honeybees.
What it does
Our system connects directly to a beehive and uses a two-tiered architecture to create a low-power yet high-performance model to detect murder hornets and other honeybee predators. Once the preliminary test detects a honey bee, a more power hungry yet comprehensive model is run and the beekeeper is notified via SMS. The beekeeper can check the dashboard to view the image of each hive and respond to incidents. The images can also be given a ground truth label, allowing the system to continue to improve through reinforcement learning.
How we built it
The preliminary test
The first test is done completely on the arduino. We use a capacitive proximity sensor to detect anything larger than an average bee. Capacitive sensors are perfect for this task as they are relatively low power, can detect non-metal objects unlike inductive sensors, and can detect thickness of objects unlike ultrasonic sensors. If the capacitance changes enough due to a large enough object getting close to the sensor, the output voltage drops which the arduino detects via a thresholding algorithm. The arduino then enables the more power hungry computer vision system.
The cloud based computer vision model
We used Google Cloud Platform's object detection system trained on a few hundred of our own labeled images to differentiate between bees and hornets. The hardware is able to upload an image after the preliminary test is triggered, allowing the computer vision model to confirm that there is a hornet or identify it as a false alarm. The beekeeper is alerted via SMS, and the incidents and images are recorded in Firebase where it can be queried by the dashboard through a NodeJS API.
The data for the dashboard is stored in Firebase and is connected through a NodeJS API. The client is built in React using a neumorphic design, as we were bored of flat & material designs and wanted a challenge. A static version of the dashboard is available at our Domain.com domain BeeSafe.Space.
Challenges we ran into
Unfortunately, we lost one brave soul in the process of building the hardware. Despite the voltage divider on input to the arduino, the output from the capacitive sensor was enough to fry our first arduino before we fixed it by adding more voltage dividers and resistors.
We implemented the neumorphism on the dashboard from scratch, which introduced challenges in maintaining code quality and sanity.
Accomplishments that we're proud of
We're proud of putting so many moving parts together into a comprehensive solution that maintains high quality hornet detection while significantly lowering power and computational needs.
What we learned
We learned what neumorphism is and how to apply it in web development, how a capacitive proximity sensor works, including how it compares to other similar sensors and how to apply it in practical projects, and how to work better as a team when integrating different parts of a project remotely.
What's next for BeeSafeSpace
We're interested in exploring different methods of neutralizing the hornet before the beekeeper arrives. One idea is to apply a small shock, stunning the hornet while not killing it to avoid killing nearby bees, then allowing the bees to swarm the hornet -- when given the opportunity, bees do fight back, and can typically take down predators by swarming them and generating heat. Another potential defense could be to automatically close or open a small door, or use a fan to blow the hornet out of the beehive.