Object Tracking via Dynamic Query Samples and Neural Feature Embedding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional tracking-by-detection methods face challenges when the number of objects in an image is not predetermined, making it difficult to effectively track multiple objects in image sequences.
Innovation Solution
An object tracking method using an artificial neural network that processes image sequences by obtaining query and detection samples from multiple frames, determining labels based on probability distributions and feature vector correlations, and employing bidirectional LSTM cells for enhanced feature extraction and reliability computation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional tracking-by-detection method is used, then object tracking can be performed in image sequences, but it becomes difficult to track when the number of objects is not predetermined
Solution Approach 1:
The system dynamically adjusts the number of query samples based on the detected number of objects in the current frame, rather than using a fixed number. This allows the tracking system to adapt to varying numbers of objects in the scene, resolving the contradiction between adaptability and complexity by making the system flexible rather than static
Solution Approach 2:
The system performs object detection on the current frame before generating query samples for the next frame. This preliminary detection action determines how many query samples to generate, enabling the system to prepare appropriately for the unknown number of objects that will need tracking in the subsequent frame
2Measurement precision
If neural network based feature embedding is used, then classification performance is improved, but computational complexity increases
Solution Approach 1:
The system extracts only the essential features needed for tracking by using neural network-based feature embedding models to convert image data into compact feature vectors. This extraction process separates the critical discriminative features from the full image data, improving classification performance while managing computational complexity through dimensionality reduction
Solution Approach 2:
The feature embedding models serve as intermediary components between the raw image data and the tracking decision-making process. These intermediaries transform complex image data into simplified feature representations that maintain classification accuracy while reducing the computational burden on subsequent tracking operations
Data Source
AI summary
Disclosed is a method for tracking an object, which is performed by a computing device including at least one processor, including: obtaining a query set including one or more query samples from a first frame included in an image sequence including two or more image frames; obtaining a detection set including one or more detection samples from a second frame included in the image sequence; and determining a label corresponding to each query sample included in the query set, based on the label of each detection sample included in the detection set.