Object Tracking Using Amodal Regions Under Partial Occlusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object tracking methods struggle to accurately track targets when they are partially occluded, leading to inefficiencies and inaccuracies in automated recognition processes.
Innovation Solution
A processor-implemented method and apparatus that utilizes a neural network model to detect amodal regions in image frames, determining search regions based on these detections, and tracks objects by applying machine learning models to template images, even when partial occlusions occur.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional object tracking methods are used, then the tracking process is simple and fast, but the tracking accuracy deteriorates when the target object is occluded
Solution Approach 1:
The patent segments the target object detection into two distinct regions: modal region (visible part) and amodal region (including occluded part). By dividing the tracking task into these segments and processing them differently, the system achieves accurate tracking during occlusion without requiring complete redesign of the entire tracking system.
Solution Approach 2:
The patent performs preliminary detection of the amodal region in the current frame before tracking occurs in the next frame. This preliminary action allows the system to anticipate the target's position even when occluded, preparing the tracking algorithm in advance rather than reacting after loss of visibility.
2Reliability
If the search region is expanded to cover the entire image frame, then the occluded target can be found, but the processing time and computational load increase
Solution Approach 1:
The patent applies local quality by making the search region adaptive rather than uniform. The search region is localized around the predicted amodal region based on motion information, applying different processing intensity to different areas: high intensity around the predicted location, low intensity elsewhere. This maintains reliability where needed while reducing overall processing time.
Solution Approach 2:
The patent uses partial action by searching only in the necessary portion of the image (around the amodal region) rather than the entire frame. This partial search provides sufficient tracking reliability under occlusion conditions while avoiding the excessive computational load of full-frame processing.
3Measurement precision
If the amodal region detection is performed using complex machine learning models, then the occlusion handling improves, but the computational resources required increase
Solution Approach 1:
The patent applies partial action by using machine learning models only for amodal region detection when occlusion is detected, rather than applying complex models continuously. This selective application maintains detection accuracy when needed while reducing overall computational energy consumption during clear visibility periods.
Solution Approach 2:
The patent changes the parameter of model complexity dynamically based on occlusion conditions. During occlusion, complex machine learning models are deployed for accurate amodal detection; during normal conditions, simpler or no model processing is used, optimizing the balance between detection accuracy and energy consumption.
Data Source
AI summary
An object tracking apparatus is provided. The object tracking apparatus includes a processor configured to detect, from a first image frame, an amodal region including a first visible region in which a target object is visible and an occlusion region in which the target object is occluded, determine, based on the detected amodal region of the first image frame, that at least a partial region of a second image frame is a search region of the second image frame, the second image frame being temporally adjacent to the first image frame, and track the target object in the second image frame based on the determined search region.