Charles Toth
Surface extraction from stereo imagery has been intensely researched and several implementations of various concepts
have been commercially available for production for many years. Although none of these techniques is flawless, the
expertise in image matching is a significant asset that can and should be exploited to complement the sparse LIDAR
spot-based surface extraction process. One of the principal problems in image matching is finding good approximate
locations. Once they are found the refinement of the matched locations is less problematic in most of the cases. In a
combined approach, the LIDAR spots can serve as initial matched locations (seed points of extremely high quality) and
then additional matching points can be searched around the LIDAR locations to densify the surface points. Various
interactions can be built into this system such as inferring from certain image patterns to an object hypothesis then
applying it to clean LIDAR data or vice versa. Fig. 4 shows an image patch with overlaid LIDAR observation and the
representative surface profiles, including LIDAR spots, stereo image-derived surface points and the
photogrammetrically determined ground truth.
190 ;
*
- * LIDAR |
ve — Topography |;
+ Photogrammetry /
186 Ka
£ i
S 184} |
i [| | .
* 1
182r i ^ +
i + he "i F
180r A i . x
> SE + ÿ / d
e = —
178 1 L SE LT ;
60 70 20 90 100-110 120. 180
Distance along the profile [m]
Figure 4. LIDAR surface elevation point locations.
The surface profiles nicely demonstrate that the LIDAR data show an excellent match to the ground truth for flat areas,
showing at the same time buildings and other man-made objects not present in the topographic surface. For example,
the first peak of the LIDAR profile from the left represents a building, while the smaller peak in the center is a car. The
stereo image-created elevation points exhibit the typical smoothed out pattern with smeared surface discontinuities (the
vertical offset noticeable over flat areas is probably due to some boresighting discrepancy). As a result of the under-
sampling of the surface by the LIDAR system, the shape of the car in the LIDAR profile is like a pyramid. Based on the
higher resolving power of the imagery, a better surface modeling of the car and its neighborhood seems feasible from
the imagery (the hypothesis for the car can be inferred from the LIDAR data and then it can be applied in the image
matching process).
5 CONCLUSIONS
In this paper we examined the feasibility of combining high-performance LIDAR data with simultaneously captured
digital images to improve the surface extraction process. The parameters used in our examples represent current state-
of-the-art LIDAR technology and commercially available digital camera systems. Our investigation was limited to the
conceptual level and addressed only a specific aspect of a rather complex topic — the question of surface sampling.
Although our discussion was incomplete, the examples, we hope, clearly demonstrated the potential of fusing LIDAR
data with simultaneously acquired imagery to improve the surface extraction process.
The already existing difference in the sampling rate between the LIDAR data and the image resolution on the ground
simultaneously offered by the digital camera provides the potential to improve the surface extraction process. Currently,
the rate is about 60 image pixels for every LIDAR spot. Since image matching on a pixel-to-pixel level is not feasible,
by assuming small clusters of pixels, a densification factor of 5 can be achieved easily even at moderate calculations
(the optimal surface point spacing vs. pixel size question itself is a topic of high interest). Since the pulse rate of LIDAR
cannot be increased without limits due to the travel time of the pulse, and the digital camera resolution is likely to grow,
thus the densification factor will continue to improve even further. This may change with the introduction of the focal
plane array LIDAR sensors.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 903
