Vision Without Limits

Traditional imaging systems consist of three main components: a CMOS sensor to capture incoming photons, an image signal processor (ISP) to convert raw data into a usable image, and GPU-accelerated routines for additional computer vision tasks such as object recognition. While effective in many scenarios, these systems have significant limitations in terms of speed, noise, and data handling capabilities.

One of the primary issues with conventional image capture is the method of averaging photons over fixed exposure intervals. This approach results in motion blur, as it captures a blend of movements within each frame. Consequently, both humans and computer vision algorithms struggle to accurately perceive dynamic scenes, reducing the overall effectiveness of these systems in critical safety applications.

To mitigate motion blur, one alternative is to shorten exposure times while maintaining the same frame rate, effectively “freezing” motion. However, conventional CMOS sensors introduce significant read noise, especially in low-light conditions where fewer photons are present. This issue is exacerbated when a high percentage of those photons are discarded, resulting in noisy images that complicate downstream computer vision tasks such as feature extraction and pattern recognition.

Ubicept addresses these challenges by utilizing Single-Photon Avalanche Diode (SPAD) sensors. Originally developed for LiDAR, they have proven valuable across other imaging applications due to their lack of read noise and ability to detect individual photons with precise timing. This capability enables clearer, more accurate images, even in low-light conditions. Furthermore, as production scales up, SPADs are becoming more cost-effective, rivaling the manufacturing costs of traditional CMOS sensors.

While SPADs provide significant advantages, they also produce an enormous amount of data. At a resolution of 1 megapixel, the data rate can exceed 100 gigabits per second. This high throughput leads to substantial power demands and extended computation times, making it impractical for conventional imaging pipeline components to manage efficiently. Addressing these bandwidth challenges is essential to fully leverage the capabilities of these remarkable sensors.

Ubicept addresses the bandwidth challenge with our Flexible Light Acquisition and Representation Engine (FLARE). FLARE manages the massive data streams from SPADs using advanced, application-specific encoding schemes that reduce data load while preserving essential information for accurate image reconstruction and analysis. The encoded photon streams are then processed by our sophisticated GPU-accelerated APIs to enable real-time data processing and analysis.

FLARE isn’t just for efficient, application-specific encodings. Its flexibility enables exceptionally precise synchronization with other sensors and illumination sources. This capability allows for advanced 3D perception, multiplexed illumination, gated imaging, and indirect sensing. And thanks to the extremely high frame rates possible with SPADs, these tasks can be performed practically simultaneously, minimizing the errors and overhead commonly associated with sensor fusion.