Machine/Computer Vision Challenge
Conventional cameras work to reconstruct images for human consumption by measuring the intensity of light on a pixel array at a fixed frame rate (typically 10-60 frames a second) . As a result these cameras have a single output mode, a continuous stream of full frame images. While this has historically been done for the accurate communication and reproduction in electronic and printed media, it makes these cameras inefficient and limiting in computer vision. This is largely due to the fact that this architecture continually captures redundant/irrelevant information while causing a combination of latency, blurring, and inaccurate signal information.
Redundant information is continuously captured even when the scene remains static. Consider the example where there is only motion in 1% of the total scene. In that scenario 99% of the pixels contain no new information yet they are continually read off of the camera and blindly processed.
Latency is incurred because of the tremendous amount of data that needs to be read off of the sensor and it must be completed within a frame time. Consider a camera that is operating at 30 fps. Waiting for the previous frame to be transferred can account for up to 33 ms of delay depending on the exposure time and that doesn't include its own data transfer or processing delays.
Blurring occurs when either an object is in motion or the camera itself is in motion (even when the scene is static). This happens because a point in the scene ends up being captured by multiple pixels during the exposure time.
While the exposure time can be shortened to limit the effects of motion blur, it negatively effects the accuracy of the signal. For example, with an exposure time of 1 ms and a fixed frame rate of 30 fps, there are now 32 ms between every frame where no data is captured. That makes the sensor blind for 97% of the time. Signal inaccuracy occurs when the sensor is no longer measuring light (looking at the scene) so it cannot capture anything happening during that time and the signal fidelity suffers.
When used in computer or machine vision, these fundamental limitations result in substantial signal inaccuracies and substantial inefficiencies in processing, power, bandwidth, and latency.
OCULI SPU™: Integrated Neuromorphic Sensing & Processing
Oculi makes the OCULI SPU™, Sensing and Processing Unit. This is the only Software-Defined Vision Sensor™ and it can deliver real-time vision intelligence (VI) by combining sensing + pre-processing in the pixel.
The OCULI SPU™ provides a disruptive architecture inspired by the eye:
Integrated sensing + processing
Parallel sensing + processing
Saliency/features (smart events) output
OCULI SPU™ is the first practical silicon that closely mimics biology in selectivity, parallel processing, and efficiency but outperforms in speed.
OCULI SPU™: Advantages
Captures high speed dynamics
High temporal resolution
Manages extreme lighting
140dB dynamic range
Wavelength and color agnostic
Demonstrated in both visible and IR
Real time image statistics at the edge
On-chip Image Signal Processor
Low latency signal to information
us vs 10's ms
Lower total size, weight, and power
mW vs 100's mW
Compatible with standard AI algorithms and general-purpose processors
Low bandwidth and low processing
Reduced bandwidth and external processing up to 99% with zero loss of relevant data.
Versatile multimodal technology
event, smart event, frame, actionable signal
Programmable features on-the-fly