SCAMP Vision System
Our system uses a vision sensor device (the “SCAMP” chip) that can carry out vision processing directly on the image sensor chip. The chip incorporates a parallel processor array, and does not output image frames, but only high-level data for further processing. This near-sensor processing provides significantly better performance and energy efficiency, as compared with state-of-the-art solutions (such as mobile GPU), while at the same time providing high temporal resolution (equivalent to >1000 frames per second).
In a conventional machine vision system, images are constantly transmitted between the camera and the processor, frame by frame, with large amount of data continually exchanged between the processors and the memory. Our approach is based on moving the computation from the processor system to the sensor itself (see Fig.2.). Essentially, we embed a parallel processor array into the image sensor. The system does not completely eliminate the need to implement a suitable external processor system, but makes it significantly simpler. The approach has a number of advantages:
- Pixel data is processed right next to where is acquired, eliminating latency and power consumption associated with the communication bottleneck between the sensor and the processor.
- Only the meaningful data, i.e. some pre-processed information extracted from the image frames, relevant to the application, is transmitted downstream for further processing (using e.g. a low power microcontroller, rather than a GPU-class processing hardware)
- Parallelism of the vision tasks is well exploited. Very large number of relatively simple processors, organised as a pixel-parallel SIMD processor array, provide large compute power and high efficiency in terms of processing power per Watt.
- Massively parallel system, with topology matching well the image data structure (2D arrays), is efficient and easy to program.
- Our unique processor circuit design, including analog and digital operations, produces extremely compact and efficient, yet fully software-programmable solution.
Figure 1. SCAMP-5 Vision System. The camera is based on a Pixel Processor Array vision sensor device.
Figure 2. SCAMP Vision Sensor. The incoming light is focused through the lens and detected by photosensors in each pixel. The image sensor array device is augmented with processing circuitry next to each pixel, implementing a SIMD processor array. This massively parallel processor array consist of Processing Elements that execute a software program “on the focal plane” of the imager. Typically only high-level processing results (e.g. the x-y coordinates of the feature locations shown in the bottom-right image), will be transmitted off chip
S.J.Carey, A.Lopich, D.R.W.Barr, B.Wang and P.Dudek, "A 100,000 fps Vision Sensor with Embedded 535 GOPS/W 256x256 SIMD Processor Array", VLSI Circuits Symposium 2013, pp.C182-C183, June 2013
Further information about the SCAMP vision sensor can be found on the SCAMP Project webpage, at the University of Manchester