Multi-Camera Target Tracking via Appearance-Based Path Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current video-surveillance systems face limitations in tracking targets across multiple cameras due to calibration errors, occlusions, and confusion between targets, especially when cameras have different fields of view, leading to poor tracking accuracy and high costs for overlapping camera setups.
Innovation Solution
A method and device for determining target trajectories using images from multiple cameras, which splits target paths into portions and generates trajectories based on appearance features, reducing reliance on location cues and correcting errors through data fusion, allowing for flexible and upgradable solutions with limited network data exchange.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple cameras with overlapping fields of view are used to increase tracking accuracy, then tracking reliability is improved, but device complexity and cost increase
Solution Approach 1:
The patent segments the tracking system into independent tracking modules, each associated with a specific camera. Each module processes images from its camera independently to generate target paths, which are then fused at a higher level. This segmentation allows the system to achieve reliable multi-camera tracking without requiring complex calibration and coordination between cameras, thus improving tracking reliability while controlling device complexity.
2Speed
If location cues are used for target tracking, then tracking speed is improved, but measurement precision deteriorates due to calibration errors
Solution Approach 1:
The patent changes the primary parameters used for tracking from location cues (which are sensitive to calibration errors) to appearance features (which are more robust). By using appearance-based tracking in independent modules and only using location information for fusion at the higher level, the system maintains fast tracking speeds while significantly improving position measurement precision and reducing sensitivity to calibration errors.
3Measurement precision
If appearance features are used for tracking, then measurement precision is improved, but tracking speed deteriorates
Solution Approach 1:
The patent applies segmentation by distributing appearance-based tracking across multiple independent tracking modules, each handling a specific camera's data. This parallel processing allows the system to maintain high precision through appearance features while preserving fast tracking speeds through concurrent processing. The computational burden is divided rather than concentrated in a single slow processing stage.
4Reliability
If data fusion is used to combine tracking results from multiple cameras, then tracking reliability is improved, but loss of information increases due to calibration errors
Solution Approach 1:
The patent extracts and removes the problematic calibration-dependent location cues from the early stages of processing. By performing appearance-based tracking independently in each module before fusion, the system takes out the unreliable calibration information that would otherwise propagate through the entire system. The fusion stage then combines only the robust appearance-based results, minimizing information loss and maximizing reliability.
Data Source
AI summary
The present invention relates to determining a trajectory of a target from two streams of images obtained from two sources of images, a sub-image of each of the images of the two streams of images representing an overlapping area of a real scene. After having obtained a target path for each of a plurality of targets, from images of the sources of images, for each of the two streams of images, each target path being obtained from a target tracker associated with a source of image, each of the obtained target paths is split into a plurality of target path portions as a function of each potential target switch along the obtained target path. Then, the trajectory is generated as a function of a plurality of the target path portions.