Related Experiment Video
Updated: Jun 3, 2026
11:18
Quantifying Learning in Young Infants: Tracking Leg Actions During a Discovery-learning Task
Published on: June 1, 2015
Trajectory learning for activity understanding: unsupervised, multilevel, and long-term adaptive approach.
Brendan Tran Morris1, Mohan Manubhai Trivedi
1Department of Electrical and Computer Engineering, University of California, San Diego, La Jolla, CA 92093-0434, USA. b1morris@ucsd.edu
Summary
This study introduces a novel framework for live video analysis, enabling real-time activity recognition and abnormality detection in surveillance systems. It effectively characterizes and predicts future behaviors using learned motion patterns.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Surveillance Systems
Background:
- Increasing adoption of video cameras generates vast amounts of visual data.
- Existing methods for managing and analyzing this data are insufficient for real-time applications.
Purpose of the Study:
- To present a framework for live video analysis.
- To enable real-time characterization, prediction, and abnormality detection of surveillance subject behaviors.
Main Methods:
- A 3-stage hierarchical learning process utilizing recurrent motion patterns and object trajectories.
- Gaussian mixture modeling for learning key activity nodes.
- Trajectory clustering for route formation and hidden Markov models for spatio-temporal dynamics.
- Online adaptation using maximum likelihood regression and periodic retraining for long-term monitoring.
Main Results:
- Demonstrated efficacy and generality of the framework across diverse datasets.
- Successful real-time characterization and prediction of activities.
- Effective detection of abnormal behaviors in surveillance footage.
Conclusions:
- The proposed framework offers a robust solution for analyzing live video streams.
- It provides effective tools for understanding and managing surveillance data.
- The approach is generalizable and performs well on various datasets.
