Abstract:
This work addresses the problem of real-time object detection in automotive environments
using monocular vision. The focus is on real-time feature detection,
tracking, depth estimation using monocular vision and finally, object detection by
fusing visual saliency and depth information.
Firstly, a novel feature detection approach is proposed for extracting stable and
dense features even in images with very low signal-to-noise ratio. This methodology
is based on image gradients, which are redefined to take account of noise as
part of their mathematical model. Each gradient is based on a vector connecting a
negative to a positive intensity centroid, where both centroids are symmetric about
the centre of the area for which the gradient is calculated. Multiple gradient vectors
define a feature with its strength being proportional to the underlying gradient
vector magnitude. The evaluation of the Dense Gradient Features (DeGraF) shows
superior performance over other contemporary detectors in terms of keypoint density,
tracking accuracy, illumination invariance, rotation invariance, noise resistance
and detection time.
The DeGraF features form the basis for two new approaches that perform dense
3D reconstruction from a single vehicle-mounted camera. The first approach tracks
DeGraF features in real-time while performing image stabilisation with minimal
computational cost. This means that despite camera vibration the algorithm can
accurately predict the real-world coordinates of each image pixel in real-time by comparing
each motion-vector to the ego-motion vector of the vehicle. The performance
of this approach has been compared to different 3D reconstruction methods in order
to determine their accuracy, depth-map density, noise-resistance and computational
complexity. The second approach proposes the use of local frequency analysis of
i
ii
gradient features for estimating relative depth. This novel method is based on the
fact that DeGraF gradients can accurately measure local image variance with subpixel
accuracy. It is shown that the local frequency by which the centroid oscillates
around the gradient window centre is proportional to the depth of each gradient
centroid in the real world. The lower computational complexity of this methodology
comes at the expense of depth map accuracy as the camera velocity increases, but
it is at least five times faster than the other evaluated approaches.
This work also proposes a novel technique for deriving visual saliency maps by
using Division of Gaussians (DIVoG). In this context, saliency maps express the
difference of each image pixel is to its surrounding pixels across multiple pyramid
levels. This approach is shown to be both fast and accurate when evaluated against
other state-of-the-art approaches. Subsequently, the saliency information is combined
with depth information to identify salient regions close to the host vehicle.
The fused map allows faster detection of high-risk areas where obstacles are likely
to exist. As a result, existing object detection algorithms, such as the Histogram of
Oriented Gradients (HOG) can execute at least five times faster.
In conclusion, through a step-wise approach computationally-expensive algorithms
have been optimised or replaced by novel methodologies to produce a fast object
detection system that is aligned to the requirements of the automotive domain.