Neural Network Object Tracking Association Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object tracking methods in videos, such as those using Kalman filters, often fail in associating object detections between frames, leading to track loss and identity switches due to erroneous associations.
Innovation Solution
A method employing a convolutional neural network to determine feature vectors that cluster object detections separately in a feature vector space, allowing for enhanced association by calculating weights based on the closeness of feature vectors between frames, thereby improving the accuracy of object detection associations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If Kalman filter based association methods are used, then tracking can be performed with simple computational methods, but association accuracy deteriorates leading to track loss and identity switches
Solution Approach 1:
The patent introduces an appearance-based weight calculation mechanism as an intermediary between the Kalman filter prediction and the final association decision. This intermediary component computes similarity weights based on appearance features (color histograms, texture, shape) and combines them with the Kalman filter probability scores, thereby improving association accuracy without completely replacing the simple computational framework
Solution Approach 2:
The patent creates a composite association score that combines two different types of information: the dynamic state-based probability from the Kalman filter and the appearance-based similarity weight. This composite approach merges heterogeneous data sources (motion dynamics and visual appearance) to achieve more reliable associations than either method could provide alone
2Measurement precision
If appearance-based weights are calculated using multiple feature vectors, then association accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent calculates appearance-based weights using multiple feature vectors (excessive action) to improve measurement precision, but applies these weights selectively rather than uniformly across all tracking scenarios. The system computes comprehensive appearance features when needed to resolve ambiguous associations, rather than continuously computing them for every frame, thus balancing precision with computational load
Solution Approach 2:
The patent pre-computes and stores appearance features (color histograms, texture descriptors, shape parameters) for each detected object in advance before the association step. This preliminary extraction of appearance characteristics allows the system to quickly compute similarity weights during association without performing complex real-time feature extraction, thereby reducing computational complexity during the critical association phase
Data Source
AI summary
A method and a device associate an object detection in a first frame with an object detection in a second frame using a convolutional neural (CNN) network trained to determine feature vectors such that object detections relating to separate objects are arranged in separate clusters. The CNN determines a reference set of feature vectors associated with the object detection in the first frame, and candidate sets of feature vectors associated with a respective one of identified areas corresponding to object detections in the second frame. A set of closest feature vectors is determined, and then measure of closeness to the reference set of feature vectors is determined for each candidate. A respective weight is determined for each object detection in the second frame. The object detection in the first frame is associated with one of the object detections in the second frame based on the assigned weights.