Adaptive Object Tracking Policy Using Reinforcement Learning

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing object tracking systems face challenges in achieving real-time performance on various computing platforms due to high computational costs of sophisticated trackers, leading to inefficiencies and inaccuracies, particularly on resource-constrained devices like smartphones, and require tedious hand-designed policies that are not scalable.

Innovation Solution

The use of reinforcement learning techniques to train an object tracking policy that dynamically selects the most appropriate tracking plan based on the computing platform and video content, optimizing between accuracy and computational resources by choosing between different object trackers and settings.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If sophisticated deep neural network object trackers are executed on every frame, then object tracking accuracy is improved, but computational cost becomes too high for real-time execution on mobile devices

Engineering Contradiction:

Improveobject tracking accuracyVSAvoidcomputational cost

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent applies dynamics by making the tracking policy adaptive and data-driven rather than static. The reinforcement learning model dynamically adjusts tracking strategies based on real-time video content analysis, computing platform characteristics, and current tracking state, allowing the system to optimize the balance between accuracy and computational cost for each specific situation

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The system changes parameters by using a reinforcement learning model to determine tracking policy parameters based on input features including video frame characteristics, object motion patterns, and platform computing power. This allows continuous optimization of the tracking strategy parameters to achieve optimal accuracy-cost tradeoff

Inventive Principle:

Principle #35Parameter changes

2Ease of manufacture

If hand-designed policies are used to select object trackers, then implementation is straightforward, but the policies are tedious to develop, not scalable across platforms, and cannot adapt to current video content

Engineering Contradiction:

Improvepolicy implementation easeVSAvoidvideo content adaptability

Core Design Contradiction:

Ease of manufactureVSAdaptability or versatility

Solution Approach 1:

The system applies self-service by using reinforcement learning to automatically learn optimal tracking policies without human intervention. The model autonomously analyzes video content, evaluates different tracking strategies, and adapts its own behavior based on observed patterns and platform characteristics, eliminating the need for manual policy design while achieving high adaptability

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The patent implements feedback mechanisms where the reinforcement learning model continuously monitors tracking performance, video content characteristics, and system resource consumption. This feedback loop enables the policy to learn from actual performance data and adjust its decisions in real-time, achieving adaptability that hand-designed policies cannot provide

Inventive Principle:

Principle #23Feedback

3Use of energy by moving object

If sophisticated object trackers are run periodically instead of continuously, then computational resources are saved, but tracking accuracy deteriorates due to gaps in detection

Engineering Contradiction:

Improvecomputational resource consumptionVSAvoidtracking accuracy

Core Design Contradiction:

Use of energy by moving objectVSMeasurement precision

Solution Approach 1:

The system applies periodic action in an intelligent way by using reinforcement learning to determine optimal sampling intervals and tracking frequencies. Rather than fixed periodic execution, the model dynamically adjusts the periodicity based on object motion speed, video content complexity, and platform resources, applying more frequent tracking when needed and reducing frequency when acceptable

Inventive Principle:

Principle #19Periodic action

Solution Approach 2:

The patent uses preliminary action by having the reinforcement learning model predict future tracking needs based on current video content analysis and object motion patterns. This allows the system to proactively schedule appropriate tracking frequencies before accuracy deteriorates, maintaining optimal performance while minimizing computational waste

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS11688077B2Adaptive object tracking policy

Publication Date: 2023.06.27 GOOGLE LLC

AI summary

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for implementing a machine-learned object tracking policy. One of the methods includes receiving a current video frame by a user device having a plurality of installed object trackers, wherein each object tracker is configured to perform a different object tracking procedure on the current video frame rent video frame. The current video frame and one or more object tracks previously generated by the one or more object trackers are provided as input to a trained policy engine that implements a reinforcement learning model to generate a particular object tracking plan. A particular object tracking plan is selected based on the output of the reinforcement learning model, and the selected object tracking plan is performed on the current video frame to generate one or more updated object tracks for the current video frame.