CNN Tracking Network Kalman Filter Bounding Box Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object tracking methods rely on low-level hand-crafted features, which are not robust to significant appearance changes and have limited discriminative power, making them ineffective in practical applications due to factors like occlusion, cluttered backgrounds, and varying illumination and viewpoints, and deep learning methods face challenges in data deficiency and inconsistency between classification and tracking tasks.
Innovation Solution
A method using a CNN with a tracking network that generates feature maps, applies operations via RPN and FCN, and utilizes a Kalman filter for bounding box estimation and adjustment, allowing for object tracking without the need for additional learning and reducing computational requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If low-level hand-crafted features are used for object tracking, then the tracking process is computationally simple, but the tracking accuracy and robustness deteriorate due to inability to capture semantic information and resistance to appearance changes
Solution Approach 1:
The patent transforms the feature representation parameters from low-level hand-crafted features to deep features extracted by CNNs. This parameter change enables the system to capture semantic information and maintain robustness against appearance changes, occlusions, and illumination variations while achieving accurate object tracking.
Solution Approach 2:
The patent replaces the traditional mechanical feature engineering approach with a data-driven deep learning approach. Instead of manually designing features, the system uses CNNs to automatically learn discriminative features from image data, substituting the manual feature extraction mechanism with an automated neural network-based feature learning mechanism.
2Reliability
If deep learning methods are applied to object tracking, then tracking accuracy improves through semantic feature capture, but data deficiency and task inconsistency problems worsen
Solution Approach 1:
The patent makes the object detector multi-functional by enabling it to perform both object detection and object tracking tasks. The same CNN architecture and learned features are utilized for both tasks, eliminating the need for separate tracking-specific training data and models. This universal approach resolves the data deficiency issue by leveraging large-scale detection datasets for tracking applications.
Solution Approach 2:
The patent merges the object detection and object tracking tasks into a unified framework. By combining these previously separate tasks, the system shares computational resources, feature extractions, and learning parameters, thereby reducing overall complexity while maintaining high accuracy for both functions.
3Reliability
If conventional CNN object detection is used for tracking, then feature representation capability improves, but computational load increases unnecessarily
Solution Approach 1:
The patent extracts and utilizes only the necessary components from the full CNN object detection pipeline for the specific purpose of object tracking. By taking out the essential feature extraction and localization capabilities while omitting unnecessary detection-specific processing steps, the system reduces computational load while maintaining effective feature representation for tracking.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for tracking an object by using a CNN including a tracking network is provided. The method includes steps of: a testing device (a) generating a feature map by using a current video frame, and instructing an RPN to generate information on proposal boxes; (b) (i) generating an estimated state vector by using a Kalman filter algorithm, generating an estimated bounding box, and determining a specific proposal box as a seed box, and (ii) instructing an FCN to apply full convolution operations to the feature map, to thereby output a position sensitive score map; (c) generating a current bounding box by referring to a regression delta and a seed box which are generated by instructing a pooling layer to pool a region, corresponding to the seed box, on the position sensitive score map, and adjusting the current bounding box by using the Kalman filter algorithm.