Many deep-learning computer vision solutions still fall short of the performance and stability needed to run in real time on low-end devices. I am travelling at the moment, and have used train journeys and breaks to improve my eye-tracking system for audiovisual performances. The initial version did its job, but with a few “quirks”.

For my current projects, I need a robust and efficient system. I explored AI implementations such as YOLO v3, v4, v8 and their “tiny” versions, as well as MobileNet-SSD v1 and v2. ChatGPT is useful here for quickly implementing and translating models between platforms; otherwise, it would have taken weeks to compare benchmarks, choose a model and implement it. I trained each model using data from a more accurate but slower version of my initial system, based on OpenCV’s k-means.

Results

• MobileNet-SSD v2: I reached 20 fps with 640×480 frames on a 2.5 GHz i7 CPU, making it the fastest.
• MobileNet-SSD v1: followed v2 closely.
• YOLO v3 tiny and others: barely reached 8 fps.

At this point, the results jittered and produced different data between almost identical frames. Reducing capabilities further to increase speed was not viable.

A change of approach

I returned to “classical” computer vision methods (soon we will call them retro), using good old reliable OpenCV. I used adaptive thresholding and median filters on a reduced image to find the region of interest (ROI), then detected contours and applied a least-squares circle fit on the original image. This achieved 40 fps on 1620×1080 images.