68% of the universe is dark matter. Even within a day, almost 12 hours have low to no visibility in many areas of the world. This is also when most of the notorious activites happen. Despite the improvement in infrared cameras and computer vision systems in visible light, there is still a gaping hole in intelligence and automation for infrared camera footage which operate in low light or hazy conditions (Visible light systems have very poor, sometimes unusable accuracy). Customers, especially in private security companies and resource constrained police officers (like those at UCPD) are in dire need of a solution that reduces the manual labor of looking at video footage, much like it has been done for the day time. However, the lack of labelled training data makes this a hard task. We wanted to use our background in Artificial Intelligence and Cloud Computing to tackle this challenge.
What it does
Our system provides Artificial Intelligence on Infrared and Thermal Imagery as a cloud computing and an on-the-edge service. Officers, Security Providers,Disaster Response agencies, Intelligence Community and Army Troops can use this intelligence on both existing or new infrastructure in order to detect, identify, track and easily search for people or objects of interest. This makes existing infrastructure smart, people more effective and reduces costs to businesses by reducing manpower needed to watch video. By making the data richer, we can also enable advanced analytics on video data, while parallely providing peace of mind. Once our solution starts operating in real time, it could provide high amount of visibility for autonomous cars as well.
Accomplishments that we're proud of
We were able to start and complete some major parts of the development of the Demo MVP to show to possible customers, Cloud Machine Learning Infrastructure and Edge Compute transformations. In addition, we are done with a considerable chunk of the training of the data augmentation tools. We were also able to verify multiple existing cloud tools and their poor accuracy on Infrared Images.
What we learned
The problem space we have chosen is very challenging, but has a very high impact in terms of business value. The lack of data also supplements the lack of research in the field, thus making it a very niche field for exploration. We got to learn about some very deep challenges in image processing and boons and banes of serverless cloud compute. We also got to learn about each other as teammates, and how we can complement each other's skills in crunch times.