Scheduling Framework for Accelerating Multiple Detection-Free Object Trackers

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Detection-free object tracking relies on Deep Neural Network (DNN) modules like Siamese networks and transformers for high accuracy.
• DNNs introduce high computational complexity, leading to slow execution speeds and bottlenecks, especially during multi-object tracking on hardware accelerators like GPUs.

Purpose of the Study:

• To develop a novel tracker scheduling framework to enhance the execution speed of DNN-based object trackers.
• To address the computational bottlenecks and delays inherent in current high-accuracy tracking methods.

Main Methods:

• Analysis of computation structures of representative trackers to derive suitable scheduling units.
• Implementation of a multi-threaded framework that decomposes and schedules tracker workloads.
• CPU-side multi-threading to optimize GPU utilization and enable parallel processing within hardware accelerators.

Main Results:

• The proposed framework improves execution speed by up to 55% for tracking two objects.
• Maintained tracking accuracy comparable to existing methods.
• Demonstrated effectiveness across various hardware accelerators, including GPUs.

Conclusions:

• The tracker scheduling framework offers a general-purpose, system-level software solution for accelerating DNN-based object tracking.
• Optimized workload management significantly enhances efficiency for multi-object tracking scenarios.