Occluded Object Tracking With Spatio-Temporal Trajectory Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object tracking algorithms struggle to accurately track occluded objects without explicit supervision, relying heavily on instantaneous observations and assuming constant object velocity.
Innovation Solution
The proposed method employs a spatio-temporal probabilistic graph to infer the trajectory of occluded objects by encoding locations of objects in a sequence of frames and using a random walk to model space-time correspondence, allowing for implicit supervision and trajectory estimation without assuming constant velocity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If conventional tracking algorithms use instantaneous observations and object permanence, then tracking can be performed with simple methods, but tracking accuracy deteriorates when objects are occluded
Solution Approach 1:
The system performs preliminary actions by encoding locations of objects in advance during visible periods, storing this information in a spatio-temporal probabilistic graph. When occlusion occurs, the pre-encoded location information and graph structure enable accurate trajectory inference without requiring complex real-time processing during occlusion events.
Solution Approach 2:
The spatio-temporal probabilistic graph serves as an intermediary structure that mediates between observed object locations and predicted trajectories during occlusion. The graph encodes spatial and temporal relationships, allowing the system to infer occluded object positions by querying the graph structure rather than directly observing objects during occlusion.
2Measurement precision
If explicit supervision is used to track occluded objects, then tracking accuracy improves, but system complexity and computational requirements increase
Solution Approach 1:
The system performs self-service by automatically encoding object locations and building the spatio-temporal probabilistic graph during training without requiring explicit supervision of occluded object trajectories. The graph structure and inference mechanisms are learned self-supervised from visible object movements, enabling accurate occluded object tracking without manual annotation of occlusion periods.
Solution Approach 2:
The system creates a probabilistic copy or representation of object trajectories through the spatio-temporal graph, which stores encoded location information and temporal relationships. This graphical copy allows the system to infer and reconstruct occluded object paths by querying the stored probabilistic representations rather than requiring direct observation or explicit supervision.
3Device complexity
If constant velocity assumption is made for occluded objects, then computational complexity is reduced, but trajectory estimation accuracy deteriorates
Solution Approach 1:
The system implements dynamics by allowing object velocities and trajectories to vary probabilistically rather than assuming constant velocity. The spatio-temporal probabilistic graph encodes dynamic movement patterns learned from observed trajectories, enabling the system to adapt velocity estimates based on object-specific motion characteristics rather than applying a fixed constant velocity assumption.
Solution Approach 2:
The system changes parameters by using probabilistic velocity distributions and temporal encoding in the spatio-temporal graph rather than fixed constant velocity parameters. The graph structure allows velocity and trajectory parameters to be inferred dynamically based on the encoded spatial-temporal relationships and object movement patterns observed during training.
Data Source
AI summary
A method for tracking occluded objects includes encoding locations of a plurality of objects in an environment, determining a target object, receiving a first end point corresponding to a position of the target object before occlusion behind an occlusion object, distributing a hypothesis between both sides of the occlusion object during occlusion from a subsequent frame of the sequence of frames, receiving a second end point corresponding to a position of the target object after emerging from occlusion from another subsequent frame of the sequence of frames, and determining a trajectory of the target object when occluded by the occlusion object by performing inferences using a spatio-temporal probabilistic graph based on the current frame and the subsequent frames of the sequence of frames. The trajectory of the target object when occluded is used as a learning model for future target objects that are occluded by the occlusion object.