Graphics Reference
In-Depth Information
In the HDR output, we are able to increase the effective dynamic range of 8- to
12-bit cameras by up to four bits (24dB), allowing simultaneous dark and light areas
in scenes to be accurately captured with little or no loss of information due to satura-
tion. We run the video camera at frame rates four times faster than the image display
rate while varying integration times from very short (to “freeze” scene motion) to
“as long as possible” (to maximize SNR). Through processing, we produce an HDR
image with significant blur reduction while simultaneously increasing the effective
dynamic range of the sensor. Furthermore, we perform contrast enhancements on the
HDR image to allow its high dynamic representation on 8-bit displays with virtually
no loss of visual information.
This coded exposure method enables high performance video to be captured from
any rapidly moving platform in real-world conditions (e.g., bright sunlight with deep
dark shadows) and displayed in low-latency real-time. It can be used in embedded
vision applications in the automotive and surveillance domains. In low light condi-
tions with significant motion blur, this system enables considerable improvements
in Detection, Recognition and Identification performance, demonstrated in real-time
in a prototype described in [ 36 ] using an FPGA.
13.5 Adaptive Task-Specific Imaging System
To conclude this chapter, we present an architectural framework for a camera sys-
tem that bridges the front-end computational imaging subsystem to the back-end
semantic reasoning. We have motivated this framework by providing two examples
of computational imaging solutions, and what is left is to make the connection to
the traditional computer vision processing that deals with detection, classification,
tracking, and learning. Our goal here is to reinforce the notion that computational
imaging functions can be made optimized for the embedded vision task at hand.
One main difference is that our framework is based on performing adaptive
feature-specific imaging directly on compressed measurements. This aspect is sim-
ilar to visual sensor networks [ 37 ] where processing is performed at the camera so
that video does not need to be transmitted. Without requiring the decoding step, we
can enable lower latency processing. In our proposed framework, we do not require
that the image be generated first, but instead, we perform some aspect of detection,
classification, tracking, and learning directly on compressive measurements. These
four components interact with each other, as shown in Fig. 13.8 . For example, track-
ing results are fed into the learning module to update models to improve detection
and classification, which in turn re-initializes tracking when tracks are lost in some
frames. In spirit, our idea is similar to that in [ 20 ] where tracking, learning, and
detection/classification are co-trained to support each other. Specifically, the four
components are implemented as follows:
Search WWH ::




Custom Search