How to distinguish between dynamic and static objects and estimate free space
Combined detection of static and dynamic objects is a data fusion challenge
Autonomous vehicles need the most precise and complete knowledge about their environment to derive driving maneuvers. Therefore, the information from multiple sensors is merged into an environment model by sensor data fusion. The environment model provides the position and motion of relevant objects as well as the drivable space around the vehicle. Since driving maneuvers are determined for some future point in time, the model should allow to make predictions for a certain period. This prediction is particularly relevant for dynamic objects like cars or cyclists. However, for a complete environment perception also static objects like curbs need to be considered to assess the free space around the vehicle, so that driving maneuvers can be properly determined.
Kalman filters and occupancy grids are multi-purpose, but not all-purpose
So far, two different approaches have been used for these two tasks – objects tracking and free space estimation. Kalman filter-based algorithms like the Extended Kalman Filter (EKF) are used to track vehicles and other road users. These algorithms use models to predict the behavior of the objects. This works very well for objects that match these models, but not if the objects behave very differently. For instance, if a system is designed to track cars, it is likely to perform insufficiently when faced with cyclists. Occupancy grid approaches are used to estimate the free space. These algorithms can detect any kind of object without the need for specific object models but has the disadvantage that it can only be used in static environments without too many missed moving objects as they usually distort the result and lead to false positive and false negative classifications.
Dynamic grid algorithm can detect static and dynamic objects to estimate free space
The Dynamic Grid is a new approach that combines both tasks in one algorithm. It is based on the PHD/MIB (probability hypothesis density/multi-instance Bernoulli) filter introduced by researchers from Ulm University and determines free space as well as static and dynamic objects in an integrated algorithm. All current automotive sensors can be used as an input to the dynamic grid: Radar, lidar, camera and ultrasound. Point clouds from lidar sensors can be used without prior clustering or pre-processing. Doppler speed measurements, as given by radar sensors, can also be considered.
The algorithm divides the environment into small areas, so-called cells. For each cell, the algorithm determines whether it is free or occupied. If it is occupied by an object, its velocity and driving direction are also calculated. Finally, objects and free areas are provided by the algorithm, e.g. for maneuver decisions and path planning. Due to its complexity, ECUs with integrated GPGPU are ideal for the use of the algorithm. The dynamic grid makes optimum use of their performance.
Exemplary use cases for the dynamic grid are:
- automated parking,
- stop & go-use cases like traffic jam pilot,
- cluttered environments with a lot of objects in the surroundings, like urban scenarios.
BASELABS development support
The BASELABS team has experiences with the implementation of the dynamic grid based on customer requirements of OEM and Tier 1 customers. Please contact us to discuss your requirements.
Contact & Downloads
In case you have any questions, please contact us.