Object Tracking via Road Plane Projection and Kalman Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for tracking moving objects, such as pedestrians, from a moving camera in a motor vehicle face challenges in accurately predicting movements and require significant computing power, as both the object and the camera are in motion, complicating the modeling of their relative movement.
Innovation Solution
The method employs an image processing system to identify and track spatial points in the road plane using a state estimator with a four-dimensional state vector, filtering out non-relevant objects and utilizing a prediction-correction structure with a Kalman filter to reduce computational effort, specifically using an unscented Kalman filter for non-linear system estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional tracking methods are used to track moving objects from a moving camera, then object tracking can be performed, but the computing power required is comparatively large
Solution Approach 1:
The patent extracts only the necessary tracking information by focusing on spatial points in the road plane rather than tracking all moving objects in three-dimensional space. This selective extraction of relevant data reduces computational load while maintaining tracking accuracy for objects of interest.
Solution Approach 2:
The tracking problem is segmented into two independent components: camera movement compensation and object movement tracking in the road plane. By separating these tasks and handling them independently, the computational complexity is reduced compared to tracking full three-dimensional movements.
2Reliability
If three-dimensional movement tracking is implemented, then comprehensive object tracking is achieved, but computing power requirements increase significantly
Solution Approach 1:
The patent transforms the tracking problem from three-dimensional space to two-dimensional road plane coordinates. By projecting all movements onto the road plane and using image coordinates, the system reduces dimensional complexity while maintaining the ability to track objects effectively.
Solution Approach 2:
The patent extracts only the essential tracking parameters (position and velocity in road plane coordinates) and discards unnecessary three-dimensional information. This extraction of minimal sufficient data reduces computational requirements while preserving tracking effectiveness.
3Reliability
If all moving objects are tracked, then no tracking information is lost, but irrelevant objects consume computing power unnecessarily
Solution Approach 1:
The patent applies different tracking strategies to different regions and object types. By focusing tracking resources on objects in the road plane (pedestrians, vehicles) and filtering out objects that do not conform to this pattern, the system optimizes computing efficiency while maintaining reliable tracking coverage for relevant objects.
Data Source
AI summary
The present invention relates to a method for the recognition and tracking of a moving object, in particular of a pedestrian, from a motor vehicle, at which a camera device is arranged. An image of the environment including picture elements is taken in the range of view of the camera device (20) by means of the camera device at regular time intervals and those picture elements are identified with the help of an image processing system which correspond to moving objects to be tracked. A picture element is extracted for each of these objects which represents a projection in image coordinates of that spatial point at which the object contacts a road plane The movement of the corresponding spatial point in the road plane is tracked by means of a state estimator which uses an at least four-dimensional state vector whose components are a position of the spatial point in the road plane and an associated speed in the road plane, wherein the tracking of the movement by the state estimator includes the steps that a prediction is generated for the state vector, this prediction is converted into image coordinates via suitable projection equations, an error to be expected for this prediction is calculated in image coordinates by means of a covariance matrix, and this prediction is compared with the picture element extracted in a later image and is updated.