PLATE: A perception-latency aware estimator.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Target tracking is crucial for numerous applications, with accuracy often increasing with computational effort (perception-latency).
• High perception-latency is not always optimal, especially when computational resources are shared or the environment demands lower latency.

Purpose of the Study:

• To develop a novel Perception-LATency aware Estimator (PLATE) for optimizing target tracking performance.
• To achieve a balance between perception-latency, detection accuracy, and computational resource utilization.

Main Methods:

• The proposed PLATE system dynamically adjusts perception configurations over time.
• It optimizes a performance measure considering both latency and accuracy trade-offs.
• Formal complexity and optimality analyses are provided, distinguishing it from heuristic frame-skipping methods.

Main Results:

• Experiments demonstrate PLATE's effectiveness in optimizing perception-latency and accuracy.
• Evaluations were conducted on a standard benchmark using real-world data.
• State-of-the-art deep learning object detection methods were utilized in the perception stage.

Conclusions:

• PLATE offers an effective approach to manage the trade-off between perception-latency and accuracy in target tracking.
• The method provides a principled way to optimize resource usage in dynamic computational environments.
• Experimental validation confirms the advantages of PLATE over existing techniques.