A motion sensor that measures what happens during the exposure

A CMOS image-sensor prototype measures how a scene changes within one exposure, encoding temporal occupancy as an analogue voltage in each pixel. The 128 × 128 device uses that information for single-frame motion estimation, rather than relying solely on the difference between two completed images. ^[1]

The appeal comes from a familiar compromise in machine vision. A longer exposure collects more light but lets a moving feature spread across more pixels. A shorter exposure preserves its shape while reducing the collected signal, potentially increasing the demand for illumination or gain.

Basler’s camera guidance treats exposure time as a motion-blur constraint and recommends keeping movement within an exposure to roughly one pixel where sharp imaging is required. That is sensible when the image itself must preserve the feature. A sensor designed to recover motion from the exposure is trying to extract different information from the same interval. ^[2]

The frame interval does not describe the blur

For an illustrative feature moving across the image at 1,000 pixels per second, a 2 ms exposure produces two pixels of travel. Extending exposure to 10 ms produces ten pixels of travel, even if the camera’s frame rate is unchanged. Those values follow directly from image-plane speed multiplied by exposure time; they are not measurements from the new chip.

Converting either result into metres per second would require the optical scale at the object. Change the lens, distance or viewing angle and the same physical motion can cover a different number of pixels. That makes a pixels-per-frame result useful for characterising a sensor, but insufficient to specify a machine’s operating speed.

The prototype divides the exposure into sub-exposure cycles and uses comparator-controlled charge injection to encode occupancy without a digital counter in every pixel. A column circuit uses a pre-event signal to resolve directional ambiguity. It supports multiplexed 10-bit imaging and motion sensing. ^[1]

Encoding the temporal information before readout moves part of the measurement into the sensor array. The output can contain more than the accumulated brightness, although the interpretation still depends on the scene and the exposure settings. A blurred outline alone does not uniquely establish how an object moved.

A different route from asynchronous event sensing

Commercial event sensors offer another way to capture change. Sony and Prophesee’s published architecture detects changes in luminance asynchronously at each pixel and outputs event position and time, rather than repeatedly transmitting an entire frame. Sony also provides event filters for flicker and unwanted events, together with control over the output event rate. ^[3]

The occupancy prototype still makes a measurement over an exposure interval, so it should not be described as an interchangeable version of that asynchronous output. The two approaches place different demands on the downstream processor. An event stream has to be interpreted over time, while an exposure-based measurement arrives with information already combined over its measurement window.

The research abstract reports detectable motion of 16 pixels per frame at 100 frames per second and power consumption of 203.16 µW in a 180 nm chip. That is sensor power, not the consumption of a complete camera and inference system. ^[1]

The published summary does not establish performance across low contrast, flickering illumination or several overlapping moving objects. Those conditions would reveal how reliably the occupancy estimate survives a less controlled scene.

Sources

• A-SSCC, author abstract for the temporal-occupancy CMOS image sensor, paper 1172
• Basler, exposure time and motion-blur guidance (background)
• Sony Semiconductor Solutions, asynchronous event-sensor operation and filtering, 9 September 2021 (background)

Image: Illustrative camera optics photograph, not the temporal-occupancy sensor prototype. Credit: Саша Пехенько / Pexels. Image source. Pexels licence.