Spatio-Temporal Tracking of Multiple Targets via DD-MCMC
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing digital video surveillance systems face difficulties in accurately tracking multiple moving targets due to occlusions, noisy segmentation, and false alarms, which violate the assumption of a one-to-one mapping between observations and targets.
Innovation Solution
The implementation of a spatio-temporal Data-Driven Markov Chain Monte Carlo (DD-MCMC) algorithm that samples the solution space to recover tracks by maximizing the consistency of motion and appearance trajectories, allowing for spatial and temporal associations between observations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional tracking algorithms assume one-to-one mapping between observations and targets, then the tracking process is simpler, but tracking accuracy deteriorates when occlusions, noisy segmentation, or false alarms occur
Solution Approach 1:
The patent segments the tracking problem into two independent components: data association (matching observations to targets) and trajectory estimation (recovering target paths). This segmentation allows the system to handle occlusions and false alarms by treating data association as a separate optimization problem, improving tracking reliability without requiring complete rethinking of the entire tracking pipeline
Solution Approach 2:
The patent introduces an intermediary data association layer that acts as a mediator between observations and targets. This intermediary component resolves the mapping ambiguity caused by occlusions and false alarms by optimizing associations based on motion and appearance models, thereby improving tracking accuracy while maintaining manageable algorithmic complexity through modular design
2Measurement precision
If the system processes all observations to recover tracks, then trajectory accuracy improves, but computational complexity increases
Solution Approach 1:
The patent segments the complex optimization problem into manageable sub-problems: motion likelihood computation, appearance likelihood computation, and their combination in the data association step. This segmentation enables precise trajectory recovery by systematically processing all observations while keeping computational complexity manageable through structured decomposition of the optimization task
Solution Approach 2:
The patent changes the parameter representation by using likelihood ratios and probabilistic models instead of deterministic matching criteria. This parameter transformation allows the system to process all observations for accurate trajectory recovery while efficiently managing computational complexity through probabilistic reasoning and optimized likelihood computations
3Reliability
If the algorithm handles occlusions and false alarms by relaxing one-to-one mapping, then tracking reliability improves, but the difficulty of detecting and measuring targets increases
Solution Approach 1:
The patent introduces an intermediary data association optimization layer that mediates between observations and targets during occlusions and false alarms. This intermediary component uses motion and appearance likelihoods to resolve ambiguous mappings, improving tracking reliability while managing data association difficulty through structured probabilistic reasoning rather than direct observation-target matching
Solution Approach 2:
The patent changes the detection and measurement approach by using likelihood-based parameters instead of direct observation matching. This parameter transformation converts the difficult data association problem under occlusions into an optimization problem based on motion and appearance probabilities, thereby improving tracking reliability while making the detection and measurement process more tractable through probabilistic models
Data Source
AI summary
Tracking multiple targets can include making different observations based on multiple different frames of one or more digital video feeds, determining an initial cover based on the observations, performing one or more modifications to the initial cover to generate a final cover, and using the final cover to track multiple targets in the one or more digital video feeds. Performing one or more modifications to generate a final cover can include selecting one or more adjustments from a group that includes temporal cover adjustments and spatial cover adjustments, and can include using likelihood information indicative of similarities in motion and appearance to distinguish different targets in the frames.