CMRT09

Stilla, Uwe

Bibliographic data

Monograph

Persistent identifier:: 856955019

Author:: Stilla, Uwe

Title:: CMRT09

Sub title:: object extraction for 3D city models, road databases, and traffic monitoring ; concepts, algorithms and evaluation ; Paris, France, September 3 - 4, 2009 ; [joint conference of ISPRS working groups III/4 and III/5]

Scope:: X, 234 Seiten

Year of publication:: 2009

Place of publication:: Lemmer

Publisher of the original:: GITC

Identifier (digital):: 856955019

Illustration:: Illustrationen, Diagramme, Karten

Language:: English

Usage licence:: Attribution 4.0 International (CC BY 4.0)

Publisher of the digital copy:: Technische Informationsbibliothek Hannover

Place of publication of the digital copy:: Hannover

Year of publication of the original:: 2016

Document type:: Monograph

Collection:: Earth sciences

Chapter

Title:: REFINING CORRECTNESS OF VEHICLE DETECTION AND TRACKING IN AERIAL IMAGE SEQUENCES BY MEANS OF VELOCITY AND TRAJECTORY EVALUATION D. Lenhart, S. Hinz

Document type:: Monograph

Structure type:: Chapter

CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation Figure 1: Detection result with false alarms. The red circle indicates redundant objects, the black mark shows objects belonging to the background. There are mainly two ways of how false alarms are being tracked (see example in Figure 1): • Collinear motion for redundant objects/features belonging to vehicles (trailer, car shadow or other). • With zero velocity if objects belong to the background (road bank, shadows of trees etc.) It is easy to see that these false alarm objects influence statistical traffic data in a manner that may lead to wrong conclusions of the traffic situation or to conflicts in model calculation. To demonstrate such influence, two examples shall be mentioned: • In a traffic scenario of a congestion where one lane moves slightly faster than the other (see Figure 2), false alarms belonging to (mainly larger) vehicles of the faster lane concurrently increase the derived density and raise the calculated average velocity. This obviously leads to a conflict in the traffic evaluation. If the false alarms belong to vehicles of the slower lane, the average velocity is lowered and thus implying an even higher vehicle density than there actually is. Therefore, it is desirable to eliminate the false alarms of the initial detection to achieve a better quality of the calculated average velocity. 3. CONCEPT OF REFINEMENT To improve the initial detection quality, we include generic knowledge about the velocity statistics and geometric layout of traffic flow in typical traffic situations (e.g. “free flowing”, “congestion”, “traffic jam”). To this end, we first track all initial detections and then eliminate the included false alarms based on an analysis of geometric layout and velocity of the trajectories. 3.1 Summary of tracking procedure Initial vehicle candidates are extracted in the neighborhood of predefined road axes using a blob detection algorithm tuned for color images. Image triplets are then used for tracking, in order to gain a certain redundancy allowing an internal evaluation of the results. A vehicle image model is created by selecting a rectangle around a particular detection. By using the shape- based matching algorithm (Steger, 2001), car hypotheses are found in the successive images. The matching procedure delivers matches in Image 2 and in Image 3. Then, new car image models are created at all hypotheses positions in Image 2 and matched to Image 3. Of course, these matches may contain multiple match results. Finally, all results obtained in Image 3 are checked for consistency including a smoothness criterion of the trajectory to determine the correct combination of the matches. A detailed explanation of this approach can be found in (Lenhart et al., 2008). The described tracking method is a very robust one delivering correct matches at about 99%, yet it tracks objects of any kind as long as their motion fulfills smoothness constraints similar to those of cars. Thus, trajectories of false initial detections are potentially tracked and also considered as “correct”. Based on the results of the tracking, the refinement is carried out. • Let us assume a snapshot of a real situation of free flowing traffic with 30 cars moving with an average velocity of about 60 km/h (which corresponds to the speed limit). By assuming 60% completeness and 70% correctness, around 18 cars will be correctly detected and there will be 8 false alarms. In case that the false alarms belong to static background they will obtain a speed 0 km/h. This leads to a calculated average velocity of 41 km/h which implies rather dense traffic and thus feeding the traffic flow model with erroneous input data. Figure 2: Congested traffic situation with different velocity in each lane. 3.2 Elimination of redundant objects A first step to eliminate false detections is to remove redundant objects from the set of detections. These are the kind of objects that belong to vehicles, such as shadows or trailers. For each pair of detections, the spatial distance is calculated. A search for very small distances delivers candidates for redundant objects. Since candidates may also include vehicles within a passing maneuver, these candidates need to be analyzed for their trajectories. The analysis includes the speed and direction of the determined trajectories and relative direction between the candidates. Identical trajectories and constant relative direction indicates redundant candidates while passing vehicles will have at least a slight difference in their speed or relative orientation. It is now tested which of the redundant candidates is the car and which is the object to be eliminated. Therefore, a quick test of the gray or color value in the center of the objects is carried out. The darker and less colored object is assumed to be the shadow and is therefore eliminated from the set of detections. In case that both objects have a similar gray or color value, the trailing object is eliminated. 182

Access restriction

Copyright

fullscreen: CMRT09

Monograph

Chapter

Table of contents

Cite and reuse

Monograph

Chapter

Image

Image fragment

Citation links

Citation recommendation