Graphics Reference
In-Depth Information
field and since this function must be smooth on image domain, the method requires
number of iterations which is computationally expensive and might not be friendly
for a real-time implementation.
Including additional information into a variational scheme can improve optical
flow results. In stereo camera systems information such as depth and vehicle ego-
motion can be used in variational approach to regularize the optical flow estimate
[
38
].
Obstacle detection involving stereovision typically uses different approaches and
various simplifications of the classic problem in order to achieve real-time perfor-
mance [
30
]. There are two main algorithm approaches depending on the domain
where calculation is performed: 3D space based and disparity based. Disparity
based algorithms are more popular as they operate directly on output from stereo
reconstruction—on disparity map. Due to limitation in resolution of the imaging
sensors and the stereo camera baseline constraints precise subpixel interpolation is
required in stereo based driver assistance camera systems.
Most of conventional stereo engines use local correlation methods that search
for pixel matches between two rectified views by comparing small patches along
the epipolar line [
41
]. Contrasting approach are global methods which exploit the
fact that the scene consist of smooth structures with very little discontinuities and
formulate the stereo problem in terms of energy function. The energy function typ-
ically includes data and smoothness terms and is then subject of optimization such
as graph-cuts, belief propagation or semi-global matching (SGM) [
42
]. The global
methods perform global disparity optimization, so chance of gross errors which
would cause unnecessary emergency breaking scenarios is minimized compared to
local correlation-based approaches which estimate disparity based on a small local
neighborhood.
SGM method calculates optimum disparity map by using dynamic programming
onmultiple one-dimensional paths crossing each pixel [
42
]. The global stereomethod
requires substantial compute andmemory bandwidth resources. First real-time imple-
mentation of SGM was published on FPGA in 2009 [
43
]. To obtain real-time SGM
performance on the CPU the key design choices are: parallelization of the most time-
consuming blocks, image subsampling, and result reuse for full resolution computa-
tion [
44
].
In some systems, the depth map is first segmented and then detected objects are
tracked over time and their velocity is estimated. In these systems, performance of
the detection heavily depends on the correctness of the segmentation. The segmenta-
tion step might merge moving object to a nearby stationary object and fail to detect a
pedestrian or cyclist moving in front of parked minivan. A solution to this challeng-
ing problem was proposed in [
45
]. The basic idea for detection is to use not only the
depthmap but also to include 3Dmotion field information. The 3Dmotion field infor-
mation is calculated by tracking points with known depth over multiple consecutive
frames.
