Fast Object Tracking on a Many-Core Neural Network Chip.

Area of Science:

• Computer Vision
• Embedded Systems
• Neuromorphic Engineering

Background:

• Fast object tracking is crucial for autonomous systems but hindered by computational complexity and hardware limitations.
• Existing solutions often lack flexibility or have not been extensively validated on diverse datasets.
• Many-core architectures offer massive parallelism and memory optimization for neural network execution.

Purpose of the Study:

• To adapt and map an attractor neural network-based object tracking model onto a many-core architecture for high-speed embedded tracking.
• To develop a hardware-friendly model with local-connection restrictions.
• To create a complete mapping framework for efficient implementation on neural network chips.

Main Methods:

• Adapted an attractor neural network model with continuous dynamics, incorporating local-connection restrictions for hardware friendliness.
• Designed a many-core neural network architecture with specific computation and transformation operations.
• Developed a mapping framework involving discretization of dynamics, a slicing scheme for topology mapping, and a constant-restricted scaling chain rule for data quantization.

Main Results:

• The adapted model achieved comparable tracking accuracy on typical video datasets.
• A many-core neural network chip was fabricated to evaluate the real execution performance.
• A single chip successfully accommodated the entire tracking model, achieving a tracking speed of nearly 800 FPS.

Conclusions:

• The proposed method enables high-speed object tracking on resource-constrained embedded devices.
• The developed many-core architecture and mapping framework are effective for deploying complex neural network models.
• This work significantly advances the capabilities of real-time intelligent monitoring and autonomous systems.