3D Model Projection Presence Map for Real-Time Object Localization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for locating objects of interest in videosurveillance, such as people or vehicles, face challenges due to geometric deformations and optical distortions, particularly in omnidirectional cameras, and suffer from high algorithmic complexity and limitations in real-time processing and background subtraction accuracy.
Innovation Solution
A method using a presence map that iteratively updates probabilities of object location based on three-dimensional models and their projections in the image space, allowing for real-time object detection by determining the overlap between model projections and background subtraction or distance images, and accounting for occultation phenomena through a constructive iterative process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If background subtraction binary mask is used to detect objects, then object detection can be performed, but the procedure exhibits high algorithmic complexity and requires significant computational resources
Solution Approach 1:
The patent segments the image space into a grid of positions and divides the object detection task into independent position evaluations. Each position in the grid is evaluated separately using 3D model projections, transforming the complex global optimization problem into multiple simpler local problems that can be processed in parallel
Solution Approach 2:
The patent transitions from 2D image space to 3D space by projecting 3D models onto the image plane. This dimensional transformation allows the system to leverage spatial relationships and depth information, enabling object detection that is robust to geometric deformations and optical distortions while reducing computational complexity through efficient projection operations
2Reliability
If background subtraction binary mask is used, then object detection is enabled, but the mask is liable to fuse disjoint groups of people or fragment normally joint groups
Solution Approach 1:
The patent introduces 3D model projections as an intermediary between the camera image and the object detection process. These projections serve as a mediator that captures spatial and geometric information, allowing the system to distinguish between objects based on their three-dimensional structure and position rather than relying solely on pixel-level background subtraction
Solution Approach 2:
The patent changes the detection parameters from binary mask values (0/1) to probability values representing the likelihood of object presence at each grid position. This parameter transformation enables the system to handle ambiguous cases more effectively, reducing false merges and splits by evaluating multiple hypotheses simultaneously
3Reliability
If shape recognition procedures are used, then object detection can be performed, but the procedures require training on labellized bases specific to particular viewpoint and camera type
Solution Approach 1:
The patent employs a universal approach by using 3D models that can represent objects from multiple viewpoints and camera configurations. The same 3D model projection mechanism works across different camera types and angles, eliminating the need for separate training bases for each viewpoint while maintaining detection accuracy through geometric consistency
4Measurement precision
If the search for people is carried out in the image space, then object location can be determined, but the procedure is not applicable for real-time processing
Solution Approach 1:
The patent performs preliminary actions by pre-computing 3D model projections for a grid of positions and storing them in advance. During real-time processing, the system only needs to evaluate which pre-computed projections match the current image, dramatically reducing the computational burden and enabling real-time object detection while maintaining location accuracy
Data Source
AI summary
In the field of videosurveillance by calibrated cameras and locating objects of interest in images, a method uses, on the one hand, an initial presence map pCP1 modeling positions i in the scene and comprising, for each position i, a value pCP1(i) representative of the probability that an object is situated at position i considered, each value pCP1(i) being obtained a location criterion defined in an image space of the image acquisition system and, on the other hand, atoms Ai predetermined for each position i of the presence map pCP1, the atom Ai of a position i comprising, for each position j, a value Ai(j) representative of overlap between projection m′(i) in the image space of a three-dimensional model M′(i) placed at the position i and projection m′(j) in the image space of a three-dimensional model M′(j) placed at position j, each three-dimensional model being representative of an object.