- Shape description by curvature (change of direction over arc length)
- recognition by aggregated edge features (smooth continuation)
- relative ego-state from features nearby (in lower part of image)
- expectation-guided, efficient feature extraction including steer angle control output.
Vehicle guidance by feed-forward control from curvature and super-imposed state feedback with integral component of lateral offset.
First demonstration of dynamic vision capabilities for road vehicle guidance to Daimler-Benz Research:
- Longitudinal & lateral guidance with lateral acceleration feedback;
- speed range from 0 to 36 km/h (10 m/s), autonomous reduction to Vmax,curve = 7 m/s circling speed with radius ~ 50 m, yielding ~ 0.1 g lateral acceleration.
- High-speed roads with multiple lanes and lane markings,
- sealed roads without lane markings, with & without curbstones,
- shadows from trees and buildings in all seasons.
- Unsealed roads with noisy transition to shoulders.
A.3.6.1a VaMoRs with 1st-generation vision system BVV2
Monocular motion stereo, an integral part of the 4-D approach; Underlying assumption:
- Gravity pulls both the obstacle and the own vehicle onto the flat ground;
- view fixation by alternative centering of gaze direction horizontally and vertically on the vertical and horizontal centers of edge features.
- Approach speed up to ~ 40 km/h.
A.3.6.1b Stop & Go behind another road vehicle (Prometheus 1991)
A.3.6.2a Cruise control with multiple vehicle detection and tracking:
Monocular stereo interpretation in own and neighboring lanes (Prometheus 1994); up to 5 vehicles in front and rear hemisphere.
A.3.6.2b Wheel recognition under oblique views:
Special region-based templates for oblique views from the side (2004)
Bifocal vision complementing an industrial radar system for more robust distance keeping, VaMP with early version of EMS-vision system;
Longitudinal control fully autonomous, lateral control by human driver to keep him „in the loop“ for better attention.
A.3.6.4 Long-lead-time studies of more demanding vision tasks – A.3.6.4a Recognition of 3-D shape and occlusion of trucks
Approximation of upper body by a rectangular box;
determine size parameters from oblique views; early hypothesis generation from partial uncovering of occlusion.
Parameterized vehicle classes from different car types to trucks,
stationary camera fixating the car driving on an oval course;
Task: Recognize vehicle type and trajectory driven.
(negative obstacles) while driving off-road on grass surface at speeds up to 10 km/h. EMS-vision including horopter-stereo with ACADIA-board (Sarnoff / PVS), saccadic perception of size of ditch for moving around it.
Collect statistical data on performance level achieved at the end of the 2nd-generation (black / white, edge-based) dynamic vision system;
> 1600 km (~ 95%) driven autonomously (longitudinal & lateral)