Valkyrie
Commercial adoption of Unmanned Aerial Vehicles is growing. Various commercial use cases, from industrial inspections to mobile security applications are springing up every day. Bringing AI technologies onboard a UAV offers unprecedented opportunities to revolutionise the commercial use of UAVs
The Problem
With increasing use of high-resolution cameras and ever growing array of other sensors, the quantity of multisensory data collected by UAVs grows exponentially. Monitoring and processing the data in real-time at the ground station requires a high-bandwidth, highly reliable always-on data connection and places a high cognitive burden on the operator flying a mission. An additional set of problems stems from the use of GPS signal for navigation – loss of signal or its unreliability under certain conditions (for instance, close to tall structures) can endanger the mission and the aircraft itself.
Our Solution
We set to create an edge AI solution capable of working on inexpensive and ubiquitous hardware with ultra-low SWAP (Size, Weight and Power) characteristics. The lower the size and weight, the higher the payload the UAV is able to carry and low power onboard AI solution does not drain the battery too quickly, prolonging the mission flight time. To prove that more hardware power is not always the key to a better AI solution we have chosen a Raspberry PI – a cheapest, commonly available system-on-chip (SOC) hardware platform, for our prototype.
To enable real-time object detection and classification in UAV video stream, our researchers built highly effective deep object detection models and developed a real-time training algorithm based on Active Learning and Self-Taught Learning principles to allow for the incremental system retraining by the user in the field to accommodate new environment or the new type of object of interest. In the process, we also effectively had to create our own flavour of 64-bit Debian operating system to ensure efficient running of our solution. Using onboard AI for object detection and classification allowed to substantially reduce the cognitive overload of an operator flying a mission, alerting her only when an object of interest appear in the camera field of vision.
To address the GPS-free navigation challenge, we have developed a deep Monocular Visual Odometry model which made possible a trajectory computation using a single onboard camera input.
The Outcome
The resulting innovative prototype UAV Edge AI solution was capable of processing onboard camera footage with 10 FPS frame rate to detect and classify objects of interest in camera’s field of vision and enable GPS-free navigation using UAV camera as input for trajectory computation. It is particularly notable that this performance was achieved on a ubiquitous, consumer grade hardware platform, weighing less than 45 grams, consuming 5V / 2.5A of power and costing less than $100
Other cases
01
Gamma
Radionuclide medical diagnostics detects how well internal organs function. The most common techniques are SPECT Tomography and Gamma Camera scans. Recording the radiation emitted by an injected radioactive tracer from within the body, they construct a 3D image – dynamic scintigram – showing how well the organs work. We used AI to analyse these images
03
Lucent
Single alkali halide crystals (boules) are indispensable in a number of a very demanding applications in high energy, nuclear, space and medical physics. Growing large crystals is an expensive and complex process with strict requirements placed on their physical properties. We used AI to improve crystallisation process quality control
04
Rura
Municipal sewers require regular inspections to ensure their proper functioning. One of the most cost-effective methods for this is using CCTV with operators controlling remotely operated crawlers which move inside the sewer pipes to collect footage. Analyzing this footage to create standardized sewer condition reports is a time-consuming and cognitively repetitive process prone to human […]
05
Porcus
Our services were requested by an innovative B2B platform for the pig growing industry. A simple yet powerful production management tool, it connects owners, caregivers and veterinarians allowing them to track the pig growing process from start to finish and keep track of animal health concerns in real time
06
Kronos
Our mission was to help an innovative company fuse wearable technology and cutting-edge computational methods to deliver personalized advice on achieving peak performance. Understanding an individual’s chronotype – their sleep and wake phases – is crucial for predicting performance and productivity throughout the day. To meet the challenge of optimizing productivity, even beyond natural chronotypes, […]
