Multi-Area Object Detection for Occlusion-Robust Person Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques struggle to accurately identify a person object across multiple frame images due to partial object impairment from varying capture angles, leading to multiple detections of the same person.
Innovation Solution
A method and apparatus that utilize bounding boxes with identification values, generating virtual bounding boxes and adjusting these values based on overlap areas to ensure consistent identification across frames, employing a Kalman filter for tracking and unifying identification values for parts of the same object.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If object detection is performed on each frame image independently, then detection speed is maintained, but the same person object is detected as multiple person objects due to partial occlusion or varying capture angles
Solution Approach 1:
The system generates virtual bounding boxes in advance for objects that may appear in future frames based on current detection results. This preliminary action allows the system to quickly match detected objects with pre-generated virtual boxes, maintaining detection speed while improving identification accuracy by preventing duplicate detections of the same object.
Solution Approach 2:
The patent introduces an identification value as an intermediary element that links multiple bounding boxes representing the same object across different frames. This intermediary allows the system to maintain separate detections for speed while using the identification value to unify them, resolving the contradiction between detection speed and identification accuracy.
2Device complexity
If multiple bounding boxes are assigned different identification values, then object tracking becomes simpler, but the system cannot recognize that multiple boxes represent the same object
Solution Approach 1:
The system continuously compares detected bounding boxes with virtual bounding boxes and updates identification values based on overlap calculations. This feedback mechanism ensures that while multiple bounding boxes may exist temporarily, their identification values are corrected to reflect that they represent the same object, preventing loss of object identity information.
Solution Approach 2:
The patent changes the identification value parameter of bounding boxes based on calculated overlap areas with virtual bounding boxes. When a detected box overlaps sufficiently with a virtual box, its identification value is updated to match, allowing the system to maintain simple tracking while preserving object identity information through dynamic parameter adjustment.
3Measurement precision
If the system generates virtual bounding boxes for all possible object positions, then object tracking accuracy improves, but computational complexity and memory usage increase
Solution Approach 1:
Instead of generating virtual bounding boxes for all possible positions uniformly, the system focuses computational resources on areas where objects are likely to appear based on current detection results and motion patterns. This local quality approach maintains high tracking accuracy in critical areas while reducing unnecessary computational complexity in areas where objects are unlikely to appear.
Data Source
AI summary
An object detection method is provided. The method includes the steps of: acquiring a plurality of frame images; detecting a first bounding box of a first object and a second bounding box of a second object from a first frame image; assigning a first identification value to the first bounding box and a second identification value to the second bounding box; generating a virtual bounding box of the first object in a second frame image; assigning the first identification value to a third bounding box of the first object detected from the second frame image, on the basis of a size of an area where the third bounding box overlaps the virtual bounding box; and changing the second identification value to the first identification value on the basis of a size of an area where the first bounding box overlaps the second bounding box.