Two-Stage Filtering for Multiple Target Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current target tracking systems face limitations in handling multiple targets, particularly in nonlinear and non-Gaussian scenarios, as they often rely on linear Gaussian assumptions and struggle with unknown and time-varying numbers of targets, leading to performance degradation and computational inefficiencies.
Innovation Solution
A two-stage filtering method using Particle Filters and density-based clustering techniques to detect and track unknown and time-varying numbers of targets, allowing for arbitrary sensor characteristics and noise distributions, and capable of handling high false contacts and missing target reflections.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If linear Gaussian tracking techniques (MLPDA, MHT, PHD) are used, then computational efficiency is maintained, but tracking performance degrades in nonlinear and non-Gaussian scenarios
Solution Approach 1:
The patent changes the fundamental parameters of the tracking system by replacing linear Gaussian models with Particle Filters that can handle nonlinear and non-Gaussian scenarios. The system uses a bank of Particle Filters with different motion models and sensor characteristics, allowing adaptation to arbitrary sensor types and noise distributions without relying on linear Gaussian assumptions.
2Reliability
If Particle Filters are used to track multiple targets, then nonlinear and non-Gaussian scenarios are handled, but the sample set tends to follow only one target
Solution Approach 1:
The patent segments the tracking problem by creating a bank of separate Particle Filters, where each filter is dedicated to tracking a single target. This segmentation allows each filter to maintain focus on one target while the overall system tracks multiple targets simultaneously. The filters work in parallel, with data association logic determining which measurements are assigned to which target.
Solution Approach 2:
The patent introduces data association logic as an intermediary component that bridges the Particle Filters and measurement updates. This intermediary layer handles the complex task of assigning measurements to the correct targets, using techniques like nearest-neighbor association and probabilistic data association to ensure that each filter receives appropriate measurements for its assigned target.
3Measurement precision
If hierarchical data fusion systems are used with PF at sensor level and Gaussian model at central level, then sensor level tracking is improved, but performance is lost during track fusion
Solution Approach 1:
The patent makes the Particle Filter universal by designing it to handle both sensor-level tracking and central-level track fusion using the same nonlinear non-Gaussian model. The same PF framework that processes raw sensor data at the sensor level is also used to fuse tracks at the central level, eliminating the need to switch to Gaussian models and preserving performance throughout the entire data fusion hierarchy.
4Adaptability or versatility
If PF-based techniques are modified to track a large number of targets with variable state vector size, then unknown number of targets can be handled, but computational cost increases significantly
Solution Approach 1:
The patent implements a dynamic filter bank architecture where the number and configuration of Particle Filters can be dynamically adjusted based on the current tracking situation. When the number of targets changes, the system can activate or deactivate filters as needed, and the state vector size of each filter remains fixed for computational efficiency. This dynamic adaptation allows handling of unknown target numbers without the exponential computational complexity of variable-state-vector approaches.
Data Source
AI summary
A two-filter based method of detecting and tracking a target that can track an unknown and time-varying number of targets, while keeping continuous track, even in scenarios with large number of false contacts or missing measurements, is provided. More specifically, a first filter provides target detection, a second filter provides target tracking of the detected targets, and a clustering technique that operates after the first filter. The first filter starts with a uniform distribution over the surveillance area and resets periodically after the clustering technique is run. When the clustering technique runs, it detects the clusters corresponding to the different targets and passes them to the second filter that tracks these targets.