International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
mentioned in chapter 3. Chapter 4 and Chapter 5 describe detail 
of algorithms. Finally, Chapter 6 is conclusion. 
3. FRAMEWORK 
From merits of Three Line Scanner Imagery for vehicle 
detection mentioned in chapter | and limitation of existing 
research, vehicle monitoring by using Three Line Scanner 
imagery has been developed under three objectives (See Figure 
2) 
To detect stopped vehicle 
To detect moving vehicle 
To classify parked and signals waiting vehicles. 
MN T 
  
| TLS forward/nadir raw images | 
  
Preprocessing 
  
| Stopped Vehicle Detection | 
| Moving Vehicle Detection _ 
i Parked/Idling Vehicle 
| Classification 
  
  
  
Figure 2 our framework of our vehicle detection algorithm 
4. PREPARATION 
Pre-processing is the preparation stage of fundamental 
information for further processing of vehicle detection. 
At first, TLS raw images are geo-coded by Chen and Shibasaki 
algorithm in [2]. Secondly, road is located and non-road surface 
are masked in TLS image in [6]. By the parallel way, due to 
many building cast shadow areas in TLS image, building cast 
shadow on TLS raw image are delineated and corrected to 
obtain ‘shadow-corrected image’ in [7]. In our study, both ‘raw 
image’ and ‘shadow-corrected image’ are region-segmented to 
generate region-segmented image and shadow-corrected-region- 
segmented image. Regions are basic unit of further processing. 
However, under-segmented regions and noise still occur on 
both images. Therefore, under-segmented regions and noise are 
corrected by erosion of morphological operation and region- 
nearest interpolation to generate 'cleaned region -segmented 
image or cleaned image shortly’ and ‘cleaned shadow-corrected 
region- segmented image or cleaned shadow-corrected image 
shortly’ respectively. At the latter step, regarding regions inside 
the road surface, all non-vehicles are rectangular polygon-fitted 
with rectangular properties such as width, length and 
length/width ratio etc. By these rectangular properties 
thresholding defined from vehicle dimension, non-vehicle 
regions are removed. Only on-street vehicle-likely regions exist 
finally. By using area-based stereo matching algorithm between 
TLS raw nadir and raw forward images, those rectangular-fitted 
polygon heights with matching correction are calculated in [5| 
Moreover, stopped and moving vehicle model hypotheses as 
explicit models are generated with generic character of stopped 
and moving vehicle in TLS single nadir images with vehicle 
dimensions under U.S. transportation law. 
  
Nadir/ Forward TLS raw images 
; Y 
| Image Geo-coding 
Road Positioning 
and non-road area masking 
| 
v = Y 
| Building Cast Shadow 
  
Image Segmentation on 
non-road masked Nadir image | Correction 
Correction of Noise | Image Segmentation 
| and Under segmentation | on Shadow-Corrected& 
= i eere ; non-road masked Nadir image 
Non-vehicle region Vehicle Dimension 
Correction of Noise | 
   
- 
Removal Database { and Under segmentation) 
= = EIE i 
| Regions Likely | Non-vehicle region 
|to be Vehicle Parts | Removal 
| on the Street Y 
Y | Regions Likely 
| to be Vehicle Parts 
Rectangular Polygons | 
| Stre 
fitted by i on the Street 
   
   
  
  
Vehicle-Likely Regions pues Y 
RUE E E es | Rectangular Polygons | 
——— re | fitted by | 
Height of 555 | Vehiele-Likely Regions | 
; Rectangular Polygons Forward/Nadir TLS TY : 
with Image processing | Height of | 
matching correlation | | Rectangular Polygons 
= 3 | with i 
p— M | matching correlation 
| Vehicle Likely Regions | Y 
| with rectangular fitted polygon Vehicle Likely Regions 
i and height | with rectangular fitted polygon 
| and height 
  
  
Figure 3 symmetric diagram of Preparation 
S. STOPPED/MOVING DETECTION 
Vehicle detection stage is the core of our study. Our algorithms 
consist of two approaches: Main approach and supplementary 
approach. Main approach is to detect moving /stopped vehicles 
automatically by using multi TLS images and to discriminate 
two classes: parked and idling vehicle class from stopped 
vehicles with on-street parking criteria. In case of omission 
from main approach, supplementary approach is to additionally 
detect vehicles from TLS single nadir image automatically and 
semi-automatically. Briefly, concepts of our vehicle detection 
approaches are mentioned as below; 
Stopped Vehicle Detection is our proposed algorithm of 
stopped vehicle detection by using multi-TLS image processing. 
At first, from Pre-processing stage some on-street vehicle-likely 
regions are selected with thresholding of height-and-matching 
correlation. This kind of thresholding is defined by selecting 
regions which are higher than road surface obtained from pre- 
processing stage. However, those regions are independent. 
Therefore pair distances among Centre of Gravities of 'those 
selected regions’ are calculated. Regarding pair distances, 
nearest regions are grouped roughly into a binary, hierarchical 
cluster tree by nearest-neighbour linkage algorithm. 
  
   
    
    
  
  
   
   
    
   
    
   
   
  
   
    
      
    
  
  
   
    
   
  
  
    
    
   
  
  
  
  
  
  
  
  
  
  
  
  
  
  
   
  
   
  
  
   
  
  
  
  
  
  
  
  
   
    
   
   
  
  
   
  
  
Interne 
  
Vehicl 
Fig