Person Tracking Using Time Windows and Virtual Nodes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing person tracking algorithms, such as the multiple hypothesis tracking (MHT) algorithm, face challenges in real-time processing, exponential calculation and memory increase, occlusion handling, and identity recognition during occlusions, leading to poor performance in video surveillance applications.
Innovation Solution
A person tracking method utilizing time windows for global tracking within each window and real-time tracking between windows, incorporating a multi-branch decision tree with virtual nodes to simulate missing frames and calculating association scores via Euclidean distance from convolutional neural network features, optimizing the tracking process for quasi-real-time results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If global tracking algorithm is used to perform person association matching according to N frames, then tracking accuracy and fault tolerance are improved, but computational complexity and processing time increase exponentially
Solution Approach 1:
The patent divides the video stream into discrete time windows of N frames each. Global tracking algorithm is applied independently within each time window rather than across the entire video sequence. This segmentation reduces the computational complexity from exponential O(2^M) for M total frames to polynomial O(N×K) where N is frames per window and K is number of windows, while maintaining tracking accuracy through continuous association across window boundaries.
Solution Approach 2:
The patent performs preliminary person detection and feature extraction for all N frames within a time window before executing the association matching algorithm. Detection results, bounding boxes, and extracted features are preprocessed and stored in advance, so that when association matching is performed, the computational burden is significantly reduced. This preliminary action separates detection complexity from tracking complexity.
2Productivity
If N frames are processed in units of time windows, then real-time processing capability is improved, but response delay increases
Solution Approach 1:
The patent makes the time window size N and sliding step dynamic rather than fixed. The parameters are adjusted based on the trade-off between real-time processing requirements and acceptable response delay. For applications requiring faster response, smaller N and larger sliding steps are used; for applications requiring higher accuracy, larger N and smaller sliding steps are applied. This dynamic adjustment optimizes the balance between processing speed and delay.
3Reliability
If virtual sub-nodes are added to simulate missing frames in the decision tree, then occlusion handling and identity recognition are improved, but device complexity increases
Solution Approach 1:
The patent introduces virtual sub-nodes as intermediary elements in the decision tree to represent missing or occluded frames. These virtual nodes act as placeholders that maintain the continuity of person identity tracking during occlusion events. Instead of breaking the tracking chain when frames are missing, the virtual nodes bridge the gap by preserving association relationships based on temporal and spatial context from detected frames, enabling robust identity recognition through occlusions.
4Measurement precision
If Euclidean distance is used to calculate person association score from CNN features, then identity recognition accuracy is improved, but computational load increases
Solution Approach 1:
The patent extracts only the most discriminative features from complete CNN feature vectors for association scoring. Instead of using all dimensions of high-dimensional CNN features, the method selects and uses only the most relevant feature dimensions that contribute to identity discrimination. This extraction reduces the dimensionality of features fed into the Euclidean distance calculation, significantly lowering computational load while preserving identity recognition accuracy.
Data Source
AI summary
A person tracking method, comprising: acquiring N frames in units of time windows; acquiring, in time windows, tracking paths of a target person according to the N frames; and constructing continuous tracking paths by means of continuous time windows, so as to obtain the tracking results of the target person.