Multi-Object Tracking via Segmented State Space Model
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The computational burden of tracking multiple objects in a video sequence becomes infeasible due to the exponential growth of the state space, making existing algorithms like the Viterbi algorithm inefficient for applications involving many objects.
Innovation Solution
A method that reduces the computational complexity by using a state space model of reduced dimensionality, defined with respect to a predefined set of states corresponding to one object, and updates an object identity map and an accumulated probability map recursively to track objects, allowing for efficient tracking of multiple objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the Viterbi algorithm is used to track multiple objects with a full state space model, then tracking accuracy is maintained, but computational burden becomes infeasible due to exponential growth of state space
Solution Approach 1:
The patent segments the joint state space of multiple objects into individual object state spaces. Instead of maintaining a full joint state space with nk states for k objects, the invention tracks each object independently using a reduced state space with n states per object. This segmentation breaks down the computationally infeasible joint probability calculations into manageable individual object tracking tasks, resolving the contradiction between tracking accuracy and computational burden.
Solution Approach 2:
The invention extracts and removes the exponential complexity component from the tracking system by eliminating the joint state space formulation. By taking out the problematic nk-dimensional state space and replacing it with n-dimensional individual state spaces, the system maintains tracking functionality while reducing computational complexity from exponential to linear growth with the number of objects.
2Reliability
If the state space model includes all possible locations for each object, then complete tracking coverage is achieved, but the dimensionality of the state space grows exponentially with the number of objects
Solution Approach 1:
The patent segments the high-dimensional joint state space into multiple low-dimensional individual state spaces. Each object is assigned its own n-state space representing possible locations, rather than maintaining a single nk-dimensional joint state space. This segmentation allows complete tracking coverage for each object while avoiding the exponential dimensionality growth that would occur in a full joint state space model.
Solution Approach 2:
The invention changes the dimensional structure of the state space from a single high-dimensional joint space (nk dimensions) to multiple low-dimensional individual spaces (n dimensions each). This dimensionality transformation allows the system to maintain comprehensive tracking coverage by preserving all possible locations for each object, while the overall computational complexity grows linearly rather than exponentially with the number of objects.
Data Source
AI summary
There is provided a method and system for tracking a plurality of objects in a sequence of images. The method comprises: receiving an image from a sequence of images; calculating, based on the received image, a detection probability map which, for each state of a predefined set of states of an object, specifies a probability that any of the plurality of objects is detected in that state in the received image; updating, based on the calculated detection probability map, an object identity map and an accumulated probability map recursively from a previous object identity map and a previous accumulated probability map corresponding to a previously received image of the sequence of images, and tracking each object in the received image based on the updated object identity map and the updated accumulated probability map.