4D Tracking via Voxel Occupancy Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer vision systems struggle to accurately recognize and track actions of people and objects in crowded and cluttered environments due to view variance, occlusions, and high computational requirements.
Innovation Solution
Utilizing multiple 3D cameras to collect depth data over time for 4D dynamic solid modeling, which constructs voxel occupancy maps to track and recognize actions by partitioning environments into partial volumes, employing machine learning algorithms for subject detection, tracking, and action recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple 3D cameras are used to collect depth data over time, then measurement precision and reliability are improved, but device complexity and cost increase
Solution Approach 1:
The patent divides the environment into discrete voxel units and further segments them into partial volumes assigned to different cameras. Each camera processes only the voxels within its field of view, breaking down the complex task of tracking all objects in the environment into smaller, manageable segments that can be processed in parallel.
Solution Approach 2:
The patent transitions from traditional 2D image processing to 3D voxel-based processing by collecting depth data from multiple 3D cameras over time. This adds the spatial dimension (z-axis) and temporal dimension (time), creating a 4D representation that improves measurement precision by capturing spatial relationships and motion trajectories that 2D systems cannot detect.
2Reliability
If voxel occupancy maps are constructed from multiple cameras over time, then tracking accuracy is improved, but computational requirements increase
Solution Approach 1:
The patent segments the voxel occupancy map construction by assigning different partial volumes to different cameras. Each camera independently processes and updates voxels within its field of view, allowing parallel computation and reducing the computational burden on any single processing unit while maintaining overall tracking accuracy.
Solution Approach 2:
The patent computes voxel occupancy only for regions where cameras actually observe objects (partial volumes), rather than processing the entire environment volume. This partial action approach reduces unnecessary computations in empty or unobserved regions while maintaining tracking accuracy in observed regions.
3Measurement precision
If depth data is collected over time for 4D modeling, then action recognition accuracy is improved, but data processing time increases
Solution Approach 1:
The patent segments temporal processing by having each camera independently update voxel occupancy for its observed partial volumes at each time point. This parallel temporal processing allows the system to incorporate historical data from multiple time points without sequentially processing all camera data, reducing overall processing time while maintaining the benefits of temporal analysis for action recognition.
Solution Approach 2:
The patent performs preliminary voxel occupancy updates for each camera's partial volume at each time point before integrating results across cameras. This preliminary action allows parallel processing of temporal data from multiple cameras, reducing the total time required to construct the 4D model compared to sequential processing.
4Measurement precision
If machine learning algorithms are employed for subject detection and tracking, then detection precision is improved, but device complexity and energy consumption increase
Solution Approach 1:
The patent applies machine learning algorithms specifically to detect and track subjects within partial volumes assigned to individual cameras, rather than applying complex algorithms to the entire scene. This local application of ML reduces computational complexity and energy consumption while maintaining detection precision in regions where subjects are actually present.
Data Source
Figure 1A
Figure 1B~1E
Figure 2A~2B
AI summary
The discussion relates to 4D tracking. One example can utilize multiple 3D cameras positioned relative to an environment to sense depth data of the environment from different viewpoints over time. The example can process the depth data to construct 3D solid volume representations of the environment, select subjects from the 3D solid volume representations, and recognize actions of the selected subjects.