Real-Time Multiobject Tracking Based on Multiway Concurrency.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Real-time Systems

Background:

• Current multiobject tracking (MOT) research primarily focuses on single-image sequences.
• Practical applications increasingly require MOT systems to process multiple concurrent video streams in parallel.
• Limited studies address the real-time performance challenges of multiway concurrent MOT systems.

Purpose of the Study:

• To propose a novel MOT framework designed for multiway concurrency and real-time performance.
• To address limitations in computing and storage resources while maintaining algorithm efficiency.
• To investigate factors influencing system real-time performance and concurrency in practical scenarios.

Main Methods:

• Developed a tracking-by-detection (TBD) based framework with expanded width and depth for parallel video processing.
• Introduced image and bounding box collectors for batch processing capabilities.
• Proposed a real-time MOT algorithm based on directly driven detection and optimized deep neural networks (DNNs) using NVIDIA TensorRT.
• Implemented a negative sample filter to ensure tracking accuracy and balance performance.

Main Results:

• The proposed framework effectively supports simultaneous processing of multiway video sequences.
• The system demonstrates good performance in handling multiple concurrent real-time video streams.
• Optimization techniques significantly accelerated DNN inference within the tracking algorithm.

Conclusions:

• The developed MOT framework offers a pragmatic solution for real-time multiway concurrency.
• The approach balances tracking accuracy with the demands of real-time processing on limited resources.
• This work advances the capabilities of MOT systems for complex, real-world video surveillance and analysis.