Object Tracking via Segmented Feature Maps and Auxiliary Bounding Boxes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning-based object tracking technologies face high computational complexity due to the extensive operations required for feature extraction and object classification and localization, which can be resource-intensive and inefficient.
Innovation Solution
The method involves generating feature maps for search and target images using depthwise convolution operations, obtaining object classification results and bounding boxes through concatenation and regression processes, and tracking objects using a weighted sum of basic and auxiliary bounding boxes, thereby reducing computational load and improving performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning-based object tracking is performed using trained neural networks for feature extraction and object classification/localization, then object tracking accuracy is improved, but computational complexity and processing time increase excessively
Solution Approach 1:
The patent divides the object tracking task into two independent sub-tasks: object classification and localization. Each sub-task uses a separate neural network model (classification network and localization network), allowing them to be processed independently and reducing the computational burden of a single comprehensive model.
Solution Approach 2:
The patent extracts and removes unnecessary computational operations from the feature extraction process. By using pre-extracted features from the search image and target image, and only performing matching operations, the method eliminates redundant feature extraction computations while maintaining tracking accuracy.
2Measurement precision
If deep learning-based object tracking is performed using trained neural networks for feature extraction and object classification/localization, then object tracking accuracy is improved, but processing time increases excessively
Solution Approach 1:
The patent performs preliminary feature extraction and prepares classification and localization features in advance. The search image feature, target image feature, classification feature, and localization feature are all pre-computed before the actual tracking decision is made, significantly reducing the processing time during real-time tracking operations.
Solution Approach 2:
By segmenting the tracking process into independent classification and localization networks that operate on pre-extracted features, the method enables parallel processing and reduces the sequential computation time required for full deep learning-based tracking.
3Measurement precision
If comprehensive feature extraction and classification operations are performed for object tracking, then tracking accuracy is improved, but computational resource consumption increases
Solution Approach 1:
The patent extracts and utilizes only the essential features needed for tracking decisions. By using pre-extracted features from the search and target images, and performing only matching operations with separate classification and localization networks, the method minimizes computational resource consumption while maintaining accurate tracking.
Solution Approach 2:
The patent performs partial feature extraction - only extracting the specific features needed for classification and localization tasks rather than comprehensive feature extraction. This selective approach reduces computational resource consumption while maintaining sufficient tracking accuracy.
Data Source
AI summary
An object tracking method includes generating a feature map of a search image and generating a feature map of a target image, obtaining an object classification result and a basic bounding box based on the feature map of the search image and the feature map of the target image, obtaining an auxiliary bounding box based on the feature map of the search image, obtaining a final bounding box based on the basic bounding box and the auxiliary bounding box, and tracking an object based on the object classification result and the final bounding box.