Pulse Detection with Minimal Resources or: How A Bottleneck Fuelled Innovation
We often get asked how our company got started—people are curious how a simple idea turned into a technology that’s now shaking things up in the industry. Like many startup stories, ours began modestly: a handful of curious minds, a stubborn problem, and a strong coffee habit. What started as a focused attempt to solve one specific challenge ended up growing into a solution with much wider applications than we ever imagined. Turns out, great things can come from just trying to fix one annoying issue really, really well.
The Humble Beginning: A Student's Challenge
A decade ago, our CIO – then he was a student – faced a deceptively simple yet technically complex challenge:
“Detect the human pulse in a face using only a video camera.”
At first glance, this problem aligns with many traditional computer vision tasks—extracting subtle visual patterns from raw image data. However, conventional approaches rely on heavy computational resources:
This task required identifying minute variations in skin color through computer vision techniques. However, the constraints were significant:
- Limited Computational Resources: Operating on an outdated, slow computer.
- Restricted Internet Access: Often working offline without cloud computing capabilities.
- Inefficient Existing Algorithms: Current methods were too resource-intensive for the available hardware.
Traditional Computer Vision Approaches
Conventional computer vision algorithms for such tasks typically include:
- Optical Flow Algorithms: Track motion between frames to detect changes but require substantial processing power.
- Fourier Transform Methods: Analyze frequency components to identify periodic signals like a pulse but can be computationally intensive.
- Deep Learning Models: Utilize neural networks to detect patterns but demand large datasets and high-performance hardware.
These methods, while effective, were impractical under the given constraints.
The Breakthrough: A Smarter, Lighter Algorithm
After months of experimentation, he developed an algorithm that was not just good enough—it was better. It worked without cloud access, on minimal hardware, and was extraordinarily fast. In fact, it could run on a simple Raspberry Pi, and video analytics were accelerated 20–1000 times compared to Python.
But how was that possible?
This wasn’t just a clever shortcut—it was a fundamental rethinking of the algorithmic foundation. He enhanced traditional Bayesian methods with a custom-built architecture that:
Uses bitwise comparisons instead of costly multiplications,
Stores training samples directly in the processor’s cache, the fastest available memory,
And executes all logic using bitwise operations, which use up to 10 times fewer processor cycles than multiplication-based calculations.
While it required highly complex, low-level programming, the result was the fastest and most efficient algorithm for such tasks ever developed under these constraints. It worked entirely offline, needed virtually no external memory access, and delivered real-time performance on edge devices.
This method has stood the test of time. It keeps delivering robust results with far fewer computational demands.
Today, this algorithm is the foundation of our solutions, mainly in biopharmaceutical manufacturing and in life sciences. What started as a student’s workaround for a resource bottleneck became a highly efficient, scalable solution that operates in demanding real-world environments.
This journey underscores an essential truth in AI and computer vision: Innovation doesn’t always require more resources—sometimes, it just requires a smarter approach.
When was the last time you turned a limitation into an opportunity?
