Lukas Bulwahn, BMW Car-IT
Michael Mühlbauer-Prassek, BMW Car-IT

Future cyber-physical systems must bring two competing properties considering their overall responsiveness together.

On the one hand, they will include algorithms the execution time of which largely depends on the cognitive complexity of the input data. This complexity cannot be measured easily in advance of executing the system's processing pipeline. For example, current algorithms for object recognition take an image as input, and detect the position of the object within that image. The required execution time depends on the algorithm's effort to detect the object with sufficient confidence. Simply speaking, object recognition in an image with background clutter takes more computing effort and time than in a clearly segmented one.

On the other hand, these future cyber-physical systems must fulfill hard real-time requirements, i.e., a late or delayed reaction can cause severe harm to humans.

A solution to this two seemingly contradictory system properties is to use adaptive computations that consider both required execution time and quality of results. If we had a system capable of balancing these characteristics by means of particular parameters, computationally demanding peak loads could be managed using less precise but still sufficient results in favor of shortened execution time.

As a first step towards building such a system, we investigate the people tracker in the collaboratively-developed, open-source Point Cloud Library. The people tracker recognizes humans in point clouds obtained from stereo cameras, and serves us as a first test subject for the question if cognitive algorithms' execution time and result quality can be adequately balanced.

In this paper, we present our evaluation of the people tracker's timing behavior. To evaluate the timing behavior, we identified parameters that we expected to have impact on the execution time. Then, we developed a test suite to provide a comprehensive insight into the correlation of execution time and result quality while varying those parameters. In order to establish real-world product-relevant conditions throughout the tests, we employed an embedded platform and a real-time capable software environment running Linux with the PREEMPT-RT patches.