Speaker
Description
Frame-based cameras read out all pixels at a fixed rate, whether or not the scene changes. In dynamic scenes, fast objects blur, dynamic range is limited, latency is bounded by the frame period, and static regions are transmitted redundantly. In event-based cameras, each pixel responds independently to log-brightness changes and asynchronously emits an event carrying its coordinates, polarity and timestamp, with latency down to 15 µs and 120 dB dynamic range, which suits fast obstacle detection and collision avoidance. However, event-based cameras and their development tools are expensive and scarce, which slows the development of event-based algorithms
We propose a low-cost converter, a board with a camera module and an STM32 microcontroller, that makes a commodity frame-based camera emulate an event-based camera, producing events (nx, ny, t, p) in real time. The microcontroller compares each new frame with a per-pixel reference log-brightness level. When the change exceeds the contrast threshold C, an ON (p = 1) or OFF (p = 0) event is emitted; a change of n thresholds yields n events and shifts the reference by nC. The timestamp t is interpolated between consecutive frames, as in ESIM and v2e, and corrected for the readout time of row ny.
An OV7670 (QVGA, 30 fps) or another DCMI-compatible sensor connects to an STM32H743. DCMI byte-select mode keeps only the luminance Y, and DMA writes it into ping-pong buffers, so frame k is processed while frame k+1 arrives. A 256-entry lookup table maps Y to log intensity, a 16-bit map holds the reference levels, and a 1 MHz timer timestamps VSYNC and HSYNC. Auto exposure, gain and white balance are disabled to avoid spurious events. Each 40-bit event holds polarity (1 bit), nx and ny (12 bits each) and time from frame start (15 bits). Events are sent via USB CDC in per-frame packets whose header carries the frame timestamp.
The key limitation is frame-based readout: a change is reported only after the next frame is captured, processed and sent. At 30 fps, waiting and readout alone take up to 67 ms, over 4,000 times longer than in an event-based camera; higher frame rates and row-wise processing shorten but cannot remove this delay. Further limitations: temporal resolution equals the frame period, with aliasing above 15 Hz; dynamic range is about 48 dB (8-bit output, fixed exposure), causing noise in dark regions; motion blur and rolling-shutter skew; and USB Full Speed cannot carry the 2.3 million events/s of full-frame changes. Thus, the converter is unsuitable for microsecond-response applications. But it has important advantages: low-cost hardware, real-time operation and output reproducing the format and sparsity of event streams. Since static regions are not transmitted, it reduces the data sent to and analyzed by the host while events per frame stay below 20% of the pixel count (a 40-bit event data packet equals five 8-bit pixels). It suits algorithm development and testing, spiking neural network training, event-based vision research and education in computer vision, robotics and automation.
| Type of presence | Presence online |
|---|