The Eye of AI #10

Computer vision is a truly transformational technology for the retail industry, but it has applications that touch nearly every aspect of our lives. This is our weekly summary of what happens when cameras, computers and great ideas get together.

Chaaarrrge!

Owning an electric vehicle is both gratifying and infuriating - all of the joy of owning a bunch of future-ready technology with the frustration of not quite having the infrastructure ready yet. Computer vision is embedded in these vehicles to help with the not-crashing and the accurately-knowing-the-speed-limit kinds of things, but a new system aims to use the same type of technology to streamline the charging process. Called 'Face-ID for cars', the idea is to use a mixture of license plate recognition and vehicle colour identification to link a car with a billing account. No more keeping a bunch of apps up to date, fiddling around with QR codes, or dealing with broken payment terminals. Neat.

Sensing Shadows

It kind of goes without saying that solar panels work best when exposed to sunlight. Of course there are many things that can get in the way and cast shadows, such as clouds, trees, or new buildings, and these have different impacts on the amount of power that can be generated. However, because of the darkness of the panel itself, and the glare of reflected sunlight, it is often difficult to tell how much of an array of panels in in the shade and therefore what the resultant hit to production will be. So researchers in China have developed a real-time shadow detection method, enhancing live video feeds by adjusting for lighting variations and identifying shadows on the modules. The system achieved high accuracy (98%) compared to other detection models and aims to improve the monitoring and maintenance of large PV arrays, offering real-time insights into shadow impacts on solar efficiency.

Never Miss a Beat

As we mentioned last week, there's a lot of new tech aimed at keeping people healthy and living in their own homes for longer. As part of that push, a team at Great Bay University have developed PhysMamba, an AI framework for measuring heart rate and physiological signals through facial videos, revolutionizing remote health monitoring. Unlike previous methods that struggled with long-range temporal dependencies, PhysMamba employs a Temporal Difference Mamba block and a dual-stream SlowFast architecture for improved accuracy. It outperforms traditional CNN and Transformer models, particularly in varying lighting conditions and facial movements. The model's success in benchmark datasets highlights its potential in noninvasive health tech, with its code available on GitHub for further exploration.

Another Lazy-AI Breakthrough

As we've mentioned before, our species has been incredibly successful because of the ability to lump things together into almost arbitrary groups which allows us to focus on only what is relevant out of the cacophony of sensory inputs. Well, this week brings another step forward in helping computers do similar things. A research team led by Prof. Wang Hongqiang from the Chinese Academy of Sciences developed a groundbreaking AI model for cross-modality machine vision. This model, called WRIM-Net, excels in extracting detailed associations across different data types, improving consistency in cross-modality image retrieval. By introducing cross-modality key-instance contrastive loss, it achieved over 90% in key performance metrics on large datasets. This innovative model has potential applications in areas like visual traceability, retrieval, and medical image analysis.

Predicting Falls for Parkinson's Patients

Researchers from Northumbria University have developed a system combining feeds from static cameras with wearable video eye-tracking glasses to improve fall risk assessment in people with Parkinson's disease (PD). While electro-mechanical inertial measurement units (IMUs) can track gait abnormalities, they lack environmental context - they don't know what the patient sees, or head-position, etc. . The new system uses computer vision to provide detailed environmental data, offering a more comprehensive understanding of fall risks. This method enhances fall prediction by integrating visual context, improving accuracy, and maintaining privacy through automatic data obfuscation.

That's everything for this week. Please keep an eye on the SAI Group blog for everything that we're thinking and talking about, including a new article this week about how the market for computer vision and Visual AI technologies is currently booming.

Got some cool tech to share? Whether its your own project, or just something you saw and thought "I want people to know about this!", let us know about it and we'll include it in upcoming editions.