Bounding Box Adjustment for Stable Object Detection Across Frames
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Mainstream object detection algorithms face instability in non-stationary images, leading to jittering bounding boxes that affect the consistency of image features and impact subsequent computer vision applications, such as face recognition and vehicle detection.
Innovation Solution
A bounding box adjustment method that compares image features between consecutive frames to adjust the bounding box selection area, using weighted averaging based on similarity thresholds to ensure smooth transitions and maintain stability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the bounding box follows the object position in non-stationary images, then the object detection capability is improved, but the bounding box stability deteriorates causing jitter
Solution Approach 1:
The patent applies preliminary action by predicting the object's position in the current frame based on historical trajectory information before performing object detection. The trajectory prediction module pre-calculates expected position, velocity, and acceleration, allowing the bounding box to follow the object smoothly without jitter while maintaining detection accuracy.
Solution Approach 2:
The patent implements feedback by continuously comparing the detected object position with the predicted trajectory position and using this information to adjust the bounding box. The system uses historical frame data and detection results to refine position estimates, creating a closed-loop control system that stabilizes the bounding box while tracking the object.
2Measurement precision
If the bounding box adjusts to object position changes, then the object tracking accuracy is improved, but the image feature consistency deteriorates
Solution Approach 1:
The patent applies dynamics by making the bounding box adjustment dynamic rather than static. The system adapts the bounding box size and position based on the object's motion state, using velocity and acceleration information to predict position changes. This dynamic adjustment maintains tracking accuracy while minimizing unnecessary bounding box changes that would cause feature inconsistency.
Solution Approach 2:
The patent uses parameter changes by modifying the bounding box parameters (position, size) based on trajectory prediction rather than direct detection results. The system changes parameters smoothly using predicted velocity and acceleration, which reduces abrupt changes and maintains image feature consistency across frames while preserving tracking accuracy.
3Loss of energy
If the bounding box stabilizes across frames, then the computational overhead is reduced, but the object detection responsiveness deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-calculating trajectory parameters (position, velocity, acceleration) from historical data before object detection. This preliminary computation allows the system to use simpler, faster detection methods in the current frame while maintaining accuracy, thereby reducing overall computational overhead without sacrificing responsiveness.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
The present application provides a bounding box adjustment method for an object detection algorithm, a device, and a storage medium. The method comprises: when object detection is performed on the basis of an object detection algorithm, acquiring a first image feature corresponding to a bounding box of the object detection algorithm in a historical frame and a second image feature corresponding to the bounding box of the object detection algorithm in a current frame, the number of frames by which the historical frame and the current frame are spaced apart being not greater than a preset frame value (S102); determining a comparison result between the first image feature and the second image feature (S104); and, using a first box selection area corresponding to the historical frame as a reference and on the basis of a degree of adjustment that is adapted to the comparison result, adjusting a second box selection area corresponding to the bounding box in the current frame (S106). According to the method, the stability of the bounding box during object detection can be improved, thereby mitigating the influence of disordered jumping of the bounding box on subsequent computer vision applications.