Occlusion-aware multi-object tracking via unsupervised re-identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current multi-object tracking systems face challenges in tracking objects across multiple frames due to occlusion, partial visibility, and the high computational cost of re-identification, especially when dealing with large datasets and numerous identities, which limits their scalability and accuracy.
Innovation Solution
The system employs an unsupervised re-identification learning mechanism that uses matching-based loss between adjacent frames, eliminating the need for identity annotation and leveraging pre-training on static image data, combined with an occlusion estimation subsystem to predict occlusion centers and estimate bounding boxes for occluded objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If re-identification is performed using traditional supervised learning methods, then tracking accuracy can be maintained, but computational cost increases significantly and scalability is limited
Solution Approach 1:
The system performs self-service by automatically generating unsupervised labels through matching-based loss computation between adjacent frames, eliminating the need for manual identity annotation. This self-labeling mechanism reduces computational overhead while maintaining tracking accuracy, as the system learns from its own observations rather than requiring external supervised data
Solution Approach 2:
The system performs preliminary pre-training on static image data before deploying to video tracking tasks. This pre-training phase initializes the re-identification model with pre-trained weights, reducing the computational burden during actual tracking and improving scalability to large datasets without requiring extensive supervised training
2Reliability
If traditional object detection methods are used, then detected objects can be tracked, but objects that are occluded or partially visible cannot be reliably detected
Solution Approach 1:
The occlusion estimation subsystem acts as an intermediary between object detection and tracking. It predicts occlusion centers and generates occlusion masks that inform the tracker about hidden objects, enabling the system to maintain reliable tracking even when objects are partially or fully occluded by other objects in the scene
Solution Approach 2:
The system replaces traditional mechanical object detection approaches with a learning-based occlusion estimation mechanism. Instead of relying on geometric or physical models to detect occluded objects, the system uses neural network-based estimation to predict occlusion centers and infer the presence of hidden objects, significantly improving detection reliability in occluded scenarios
Data Source
AI summary
A system for tracking a target object across a plurality of image frames. The system comprises a logic machine and a storage machine. The storage machine holds instructions executable by the logic machine to calculate a trajectory for the target object over one or more previous frames occurring before a target frame. Responsive to assessing no detection of the target object in the target frame, the instructions are executable to predict an estimated region for the target object based on the trajectory, predict an occlusion center based on a set of candidate occluding locations for a set of other objects within a threshold distance of the estimated region, each location of the set of candidate occluding locations overlapping with the estimated region, and automatically estimate a bounding box for the target object in the target frame based on the occlusion center.