Time-Series Image Selection for Adaptive Neural Object Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object tracking techniques using multilayered neural networks face challenges in training robust models due to the need for large datasets and the difficulty in adapting to various tracking target subjects, especially when devices have limited computational resources and processing times.
Innovation Solution
An information processing apparatus that selects reference and search images from time-series images based on capture time and dissimilarity, updates neural network parameters using inference results and ground truth data to adapt to varying target subjects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If the number of parameters of a multilayered neural network is drastically reduced to fit limited device resources, then the device can perform object tracking, but it becomes difficult to train a model adapted to all various tracking target subjects
Solution Approach 1:
The patent changes the parameters of the neural network by selectively updating only specific parameters (those related to appearance variation) while keeping other parameters fixed, allowing the model to adapt to various targets without requiring a complete retraining of all parameters
Solution Approach 2:
The patent segments the parameter update process into two distinct phases: first freezing all parameters during normal operation, then selectively updating only appearance-related parameters when appearance variation is detected, rather than updating all parameters uniformly
2Measurement precision
If a multilayered neural network with many parameters is used to achieve accurate object tracking, then tracking accuracy improves, but devices with limited computational resources and processing times cannot support it
Solution Approach 1:
The patent applies partial action by updating only the necessary appearance-related parameters rather than all parameters, achieving sufficient tracking accuracy while reducing computational burden on resource-constrained devices
Solution Approach 2:
The patent implements dynamic parameter updating where the system adaptively decides whether to update parameters based on detected appearance variation, allowing the model to maintain high accuracy when needed while conserving computational resources during normal operation
3Reliability
If independently prepared training datasets are created for each device function, then the model adapts well to that specific function, but it cannot handle variations in target subjects effectively
Solution Approach 1:
The patent creates a universal parameter update mechanism that can handle appearance variations across different target subjects and device functions, making the trained model more versatile without sacrificing function-specific performance
Data Source
AI summary
An information processing apparatus includes at least one processor and at least one memory. The at least one memory stores instructions for causing the at least one processor and the at least one memory to obtain a plurality of time-series images; select a reference image and a search image from among the plurality of time-series images based on at least any of a time at which the plurality of time-series images is captured, a predetermined time interval, and a dissimilarity degree between the plurality of time-series images; and infer, based on the reference image and the search image selected from among the plurality of time-series images, a target subject in the search image that corresponds to a target subject in the reference image to update a parameter of a neural network based on an inference result and ground truth data.