International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
  
  
  
  
  
  
  
  
  
  
  
Flight 1 Flight 2 
Flying Height (AGL): 470 m 660 m 
Average Ground Speed: 56.6 m/sec 58.5 m/sec 
Heading: 290 degrees 250 degrees 
Scan Frequency: 50 Hz 46 Hz 
Field of View (Half 6 degrees 20 degrees 
Angle): 
Laser Repetition Rate: 10 kHz 70 kHz 
Point density 1.5 points/m* 2.4 points/m” 
Area Route 35, OH Toronto, Canada 
  
  
  
  
Table 2. Flight parameters 
Practically, in the Ohio dataset, 20-30 points were collected 
from a passenger car (and 40-60 points from a truck) traveling 
along the direction of flight. In the opposite direction, the value 
of collected points/vehicle is under 10, based on which we were 
not able to reliably determine the vehicle profile. However, 
from at least 20 points, reasonable vehicle profiles could be 
calculated 
Due to multipath reflection or processing errors, the coordinates 
of some points are wrong, but our main problem is caused by 
out-of-shape points. For classification, we need points 
describing the profile of the vehicle; therefore, airborne LiDAR 
seems to be an appropriate technology for this purpose, since 
points are reflected from the top of the vehicle. However, the 
flight line does not always coincide with the road (or centerline 
of the road) and from the side of the scan swath points tend to 
reflect also from the side of vehicles; these points have smaller 
height values than the profile points. 
Points reflected from the side of vehicles, or from roadside 
objects, as well as multipath reflection can corrupt the results. 
In order to eliminate this effect, the sides of vehicles have been 
cropped; only points in the middle strip were used for creating 
the profile (Lovas 2004b). 
Shape Detection 
First, the vehicles are divided into equal parts along their 
profiles (not laser profile, which is across!), each containing at 
least two points. In every section, the mean height value of the 
points, ie, the points of the profile are calculated. Our 
experiments show that the minimum resolution should be at 
least 8 sections along the profile: this makes also visual 
recognition possible. In case of a passenger car, the engine hood 
and the trunk lid are clearly shown, so is the MPV's bulky 
figure (presumably an SUV or van), and the shape of the 
eighteen-wheeler is even better outlined on the picture below 
(Figure 3). 
Passenger car Multi purpose vehicle 
NUUED NO ey 
Truck 
Figure 3. Typical vehicle shapes 
The partition of each vehicle depends on the number of 
reflected points and the point distribution. All the 
vehicles used in this shape-definition were previously 
classified with PCA-based clustering and controlled by 
visual recognition based on color-coded 3D models. 
In order to compare the shapes, we have to normalize for 
length. The deviation in actual vehicle length is only one factor 
(more than 90% of vehicles are between 4 and 5 meters); the 
other contributor is elongation which is proportional to speed. 
The previously categorized vehicles tend to have similar 
profiles (Figure 8). Creating a buffer zone (envelope curve) 
based on profile points with extreme values shows the key to 
the profile based vehicle detection. These buffer zones have to 
be created for each predefined category; every new vehicle 
profile is evaluated whether it fits or not in the particular buffer 
zone. The upper and lower boundaries are marked in red, the 
sample shape in green, respectively. 
4. GROUND-BASED LIDAR DATA 
In general, ground-based laser scanning technology has similar 
advantages as airborne LiDAR. It rapidly produces high 
density, accurate spatial data. On the ground, the sensor 
coordinates can be obtained by accurate positioning methods, 
such as DGPS or conventional geodetic measurements. 
Recently LiDAR sensors are frequently complemented with 
digital camera systems, and besides the 3D coordinates and the 
laser intensity values, the “color” of each point, can be attached 
(Figure 4). 
  
  
Figure 4. Test cars (laser points fused with camera image) 
The main difference lies in the point density; ground laser 
sensors usually mounted on tripods, the objects to be mapped 
are not moving during measurements, and hence there is 
enough time to scan through the arca with the desired point 
density. Of course, the application area of airborne LiDAR 
significantly differs from that of ground sensors. similarly to the 
relation between airborne and close-range photogrammetry. 
In our first approach, we derived the “mean-shape” from the 
training data set and then generated a buffer-zone for the 
classifier. In order to improve the resolution of the profile, we 
performed a ground-based laser scanning campaign in a packed 
Interna 
parking 
vehicle 
Catega 
The gr 
in Hu 
represe 
airborn 
charact 
ninetie: 
sold in 
MPVs 
SUVs : 
proport 
proport 
we foc 
regions 
  
For ou 
acquire 
profile, 
popular 
PCA Ti 
First, th 
cars ha 
paramet 
are on t 
and aga 
are eith 
direction 
  
  
Fi