208
GROUND PLANE OBJECT TRACKING UNDER EGOMOTION
Wlodzimierz Kasprzak
Bavarian Research Center for Knowledge Based Systems
FORWISS, Am Weichselgarten 7, D-91058 Erlangen
E-mail: kasprzak@forwiss.uni-erlangen.de
Tel. +49 9131 691-194, Fax -185
KEY WORDS: Image Sequences, Kalman Filter, Motion Estimation, Object Recognition.
ABSTRACT:
An adaptive approach to moving object recognition in image sequences from a single low-cost camera under egomotion
is proposed. The object motion is restricted by ground plane trajectories. Quantitative and qualitative results of vehicle
tracking in traffic scenes are presented. It is demonstrated that acceptable results with a single low cost camera can
be achieved for those objects, which are projected to image regions not smaller than 10x12 pixel”.
1. INTRODUCTION
Recently more intensive work on object tracking in traffic scenes has been done, but mostly limited to a stationary
camera case (without egomotion) (Koller, 93), (Tan, 93). The problem of vision systems for navigation purposes, as
addressed for example in (Masaki, 92), consists of the recognition of road boundaries and moving obstacles within
the road area. The detection of obstacles, that are very distant from the camera and a precise estimation of their
positions, motions and orientations on the road plane is still a challenging problem (Regensburger, 94). Additionally to
the non-parallel projection, there is a problem of stable camera orientation detection (overcoming the camera nidding
movement) and of recognizing the road (if many obstacles exist in the scene). These are reasons why standard low
level motion estimation (Schunck, 81) fails to provide us with discriminant object features in this case, even if the
translational and circular velocities of the camera vehicle (egomotion) are known.
In this paper a model based approach to moving object recognition in image sequences under egomotion is described.
À specific application in mind is a road scene analysis system (Kasprzak, 94). Due to large discretization errors in
images of outdoor scenes, the scene and object domains are restricted by shape assumptions and the object motion
trajectories are restricted to a ground plane circular movement. In section two an overview of the adaptive object
recognition approach is given. The main computational steps of this approach - initialization and single object tracking
- are discussed in section 3. Test results of moving object recognition in monocular image sequences of traffic scenes
and a summary follow in subsequent sections.
2. THE RECOGNITION-BY-TRACKING APPROACH
2.1 Adaptive object recognition
There are at least two different approaches to the recognition of moving objects in image sequences. The first approach
tries to initialize a nearly correct hypothesis, spending a lot of time on the estimation of an application-independent
visual motion — either the optical flow is computed and segmented afterwards (Koller, 93) or discrete features are
tracked and their image motion is interpreted in terms of object's motion and depth (Kasprzak, 93). This type of
initialization corresponds to hypothesis dependent measurements, i.e. the judgement of new measurement is directly
dependent o how well it fits the hypothesized object. This approach can be called recognize-and-track.
Here a different approach to model-based object recognition in image sequences is proposed. As the motions of
tracked objects are not significantly different from the egomotion and the nidding of the camera is disturbing the
estimated visual motion, a geometry and model-based object initialization is performed with default motion. No
precomputation of visual motion is necessary. Accordingly, the judgements of consecutive measurements are not
dependent on the predicted hypothesis, but on how well they fit the global model expectations. Due to this statistical
independence the Kalman filter (Wünsche, 88) for adaptive estimation of the tracked hypothesis can be applied. In
this way the whole hypothesis can largely be modified during the consecutive analysis and the approach can be called
recognition-by-tracking. :
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences”, Zurich, March 22-24 1995