In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3A/V4 — Paris, France, 3-4 September, 2009 
39 
the samples S'” eS r 1 <m< n Y . Once the principal geodesics are 
available for each C y , the classification of an unlabeled sample 
x can be performed by finding the category with the closest first 
principal geodesics to x. The corresponding motion status of a 
vehicle is found by 
/ =argmin||log(//' l v(11 x)||, ye {1,2} (6) 
Generally, it is claimed that the classification of vehicle status 
can successfully run based solely on the first principal 
geodesics of a movement category. Although there are 
significant variations in shape over one category, the first 
principal geodesics H is assumed to summarize the essential 
shape features of vehicle point sets in terms of only 
distinguishing between binary motion statuses. 
3.3 Results 
We used the same vehicle datasets as derived in the section 2 to 
assess the proposed algorithm intended for classifying the 
motion status. Both of datasets are acquired over 
300x400 m 2 dense urban areas with averaged point density of 
about 1.4 pts/ m~ . The only one difference between them is that 
the first one used is co-registered from multiple strips rather 
than one-path. The classification results of vehicle motion status 
are presented in Fig.5. To access the performance of Lie group 
based classifier, minimum distance classifier was used to 
classify the same datasets based on the feature space spanned 
by vehicle parametrization. 
The test dataset each consists of more than 50 vehicles 
successfully detected by vehicle extraction process. A set of 5 
vehicle samples from each motion category is manually 
selected to train the classifier for vehicle motion status at first. 
It can be expected that poorly chosen training samples due to 
the strong shape variability in the category of moving vehicle 
could have a negative effect on classification performance. 
Therefore, the selection of training data for moving vehicle 
category should be carried out in such way that the fundamental 
shape information are expressed and generalized. Receiver 
Operating Characteristic (ROC) curves are generated by 
comparing classification results with reference data manually 
acquired by human interpretation and shown in Fig.6 for 
respective test datasets. 
'atf 
~ w 
(b) 
Figure 5. Vehicles motion classification results for dataset 1 and 
II (top-view of vehicle point sets). Blue: moving; Red: 
stationary; Yellow: uncertain. 
(a) (b) 
Figure 6. ROC curves for vehicle motion classification, (a) 
Dataset I; (b) Dataset II. 
3.4 Discussion 
Since we do not have real “ground truth” for vehicle motion 
which could be simultaneously captured along the scanning 
campaigns by an imaging sensor as described in Toth and 
Grejner-Brzezinska, (2006), the results are firstly assessed with 
respect to human examination abilities. Based on the context 
relations the vehicle movement could be roughly distinguished 
between moving vehicles and stationary ones. Note that the 
along-track motion cannot be resolved on principle if the true 
length is unknown, our evaluation are inherently biased by 
ambiguities introduced by the incorrect vehicle length. 
It can be found out from the results displayed above that most 
of detected moving vehicles appear in the heavily travelled 
roads such as flyovers and main streets of city and the vehicles 
classified as motionless are mostly found in the parking lots or 
along road margins. The yellow class indicates the vehicles of 
uncertain status which are all nearly placed very close to each 
other in a parking lot and are excluded from the motion 
classification step due to the shape irregularity. False alarms 
from motion classification by our approach usually appear for 
slowly moving vehicles which travelled not perpendicular to the 
flight direction or those moving ones that are shaped by 
anomaly sample points in ALS data due to vegetation occlusion 
or unstable reflection properties. As indicated in ROC curves, 
the Lie group based classifier outperforms the minimum 
distance classifier in both cases, as its ability to generalize 
various shapes from training data, even for worst-cases, is 
demonstrated. It can also be observed that the second test