Object Tracking Filters Using Temporal Decorrelation of Detections
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object detection algorithms based on deep learning suffer from temporal correlation of detections, violating the foundational assumptions of efficient estimators and leading to inaccurate tracking and control in autonomous systems.
Innovation Solution
Implementing a processor-based system that selects non-overlapping sets of sensor data observations for each filter to generate a track, using filters like Kalman filters, ensuring each frame of data is used only once, and controlling autonomous robots based on these tracks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If object detection algorithms operate on a temporal window of sensor data, then detection accuracy is improved, but temporal correlation of detections occurs which violates estimator assumptions
Solution Approach 1:
The patent segments the temporal window of sensor data into multiple non-overlapping sets, where each set is assigned to a different filter instance. This segmentation eliminates temporal correlation between detections processed by the same filter while maintaining the benefits of temporal window processing for detection accuracy.
Solution Approach 2:
The patent introduces a new dimension of filter instance multiplication to resolve the temporal correlation problem. Instead of processing all temporal data through a single filter (one-dimensional processing), multiple filter instances process different segments of temporal data in parallel, transforming the problem from a single temporal stream to a multi-instance parallel processing architecture.
2Reliability
If multiple filter instances are used to process different sets of observations, then temporal correlation is reduced, but system complexity increases
Solution Approach 1:
The system segments both the input data and the filter instances in a coordinated manner. Each filter instance receives a specific segment of observations and produces decorrelated detections. This segmentation approach achieves temporal decorrelation while managing complexity through systematic data division rather than arbitrary complexity increases.
Solution Approach 2:
The patent creates multiple copies of the filter instance, each configured to process a specific non-overlapping set of observations. This copying approach simplifies the system architecture compared to designing entirely different processing paths, as each filter copy can be an identical template instantiated multiple times with different data inputs.
3Reliability
If sensor data from each frame is used only once by filters, then temporal decorrelation is achieved, but data utilization efficiency decreases
Solution Approach 1:
The patent segments the temporal data stream into non-overlapping sets that are distributed to different filter instances. Each filter instance uses its assigned data exactly once, achieving temporal decorrelation. The segmentation ensures complete data utilization across all filter instances, with no data wasted or redundantly processed by a single filter.
Solution Approach 2:
The patent merges the outputs of multiple filter instances that each processed different segments of the temporal data. This merging combines the decorrelated detections from all filter instances into a unified detection stream, achieving both temporal decorrelation and efficient utilization of all sensor data across the complete observation window.
Data Source
AI summary
Systems and methods for tracking an object. The method comprising: receiving, by a processor, a series of observations made over time for the object; selecting, by the processor, a plurality of sets of observations using the series of observations; causing, by the processor, the plurality of sets of observations to be used by at least one filter to generate a track for the object (wherein the at least one filter uses sensor data associated with each of a plurality of frames of sensor data only once during generation of the track); and causing, by the processor, operations of an autonomous robot to be controlled based on the track for the object.