In our last post, we spoke about AI and how FPGAs help enables their development. Today we drill down into a smaller subset - Computer Vision (CV). Computer vision is the field of AI that focuses on enabling computers to identify and process objects in images and videos. Thanks to advances in deep learning, it is now possible for systems to reach accuracy rates of 99% for object recognition. There is a huge market for CV, with the most common applications being in self-driving cars, healthcare, and facial recognition.
Another interesting aspect to consider is smart surveillance. Surveillance is an important but sometimes overlooked industry, with products like Amazon’s Ring doorbell being valued at $1 billion and the global video surveillance market is forecast to be $87.36 billion by 2025.
Surveillance is so integral to our society as it automates protection and is a form of active prevention. Taking construction as an example: Surveillance is one of the most important risk mitigation measures at construction sites, for both safety and security reasons.
A worksite without surveillance is like a house without locks. Not everyone will attempt to enter, but those with ill intentions would be glad for the lack of resistance. Securing a worksite isn’t easy, as traditional CCTV systems and alarms are costly and time-consuming to configure, especially on short-term projects or unpowered locations. Because equipment and raw materials are easy to sell and difficult to trace, construction companies face a constant problem of how to secure their work sites 24/7.
Beyond the prevention of theft, surveillance can also detect potential health and safety issues to prevent injuries and accidents. It is estimated that 60,000 incidents occur annually within construction sites worldwide, with accidents delaying large projects by an average of 20%. Spotting issues like materials being stored precariously or workers not utilizing safety gear can prevent unnecessary accidents.
What can a ‘smart’ surveillance system do?
A system can become smart by becoming capable of analysis and action. The goal of a smart surveillance system is to create a network of lightweight cameras providing data for real-time analysis utilizing computer vision. The advances in CV mean that a smart system has the potential to detect various forms of activity and objects. From individuals not wearing protective equipment, to tracking materials and workers to even detecting unauthorized personnel. It is also a good complementary tool to the existing workforce, being capable of improving their effectiveness instead of going with the approach of simply hiring more guards or supervisors to patrol the grounds.
For equipment storage (identification of designated areas for storage)
However, one obstacle remains, which is to do it locally. Current AI solutions to analyze site conditions require data to be passed on to their engine over the cloud. This means that if the dataset to be sent over for analysis is large, the bandwidth might not be able to support the load, resulting in a delay in analytics output. In addition, for data to be sent to the cloud for processing, a data collection gateway needs to be connected to the internet.
As most sites tend not to have the fully functional infrastructure and power, locations with limited internet connectivity can see attempts at data transmission and processing being severely handicapped. Power constraints also mean that any on-site processing equipment has to be efficient and not a power sink.
How FPGAs enable smart surveillance
The inclusion of FPGAs in a surveillance system enables efficient on-site computation and processing, removing the need for data to be sent to the cloud and so removing the reliance on internet connectivity. An FPGA based system would be able to support some of the most popular analytics frameworks such as TensorFlow and PyTorch for features including but not limited to accurate object recognition and material tracking at construction sites.
This combined with the power efficiency of FPGAs addresses the unique requirements needed for deployment at construction sites such as power constraints and limited internet connectivity while allowing for the flexible implementation of computer vision algorithms. Users can switch between algorithms, depending on the needs at the time and additional features can also be added on as computer vision algorithms advance.
This flexibility is key to creating a system that can scale and upgrade itself as AI frameworks improve. A system that can remain as relevant as the latest CV algorithms will allow users to unleash the full power of AI.
If you’d like to discuss or learn more about such systems, please reach out to us here.