Study on the Moving Target Tracking Based on Vision DSP.

Area of Science:

• Computer Vision
• Embedded Systems
• Digital Signal Processing

Background:

• Embedded visual tracking demands high real-time performance and system resources, posing challenges for systems with limited hardware.
• Optimizing hardware resource utilization is crucial for efficient visual tracking algorithms.

Purpose of the Study:

• To evaluate hardware optimization methods for embedded visual tracking systems.
• To investigate the real-time performance of a kernel correlation filter (KCF) algorithm on a vision digital signal processor (DSP).

Main Methods:

• Implemented and optimized a KCF tracking algorithm on a vision DSP.
• Analyzed the impact of data parallelism (DP), instruction parallelism (IP), DSP core parallel processing, and integrated direct memory access (iDMA).
• Introduced a scale filter to address KCF's limitations with scale transformation and employed a time-sharing strategy to boost system speed.

Main Results:

• The optimized KCF algorithm demonstrated efficient utilization of limited hardware resources.
• Real-time tracking speed and system resource usage met expected requirements.
• The KCF algorithm with a scale filter achieved accuracy comparable to the discriminative scale space tracking (DSST) algorithm.

Conclusions:

• Hardware optimization techniques effectively improve the performance of embedded visual tracking systems.
• The KCF algorithm, enhanced with a scale filter and optimized for vision DSP, offers a viable solution for real-time visual tracking with limited resources.
• The findings are applicable to other machine vision algorithms requiring efficient hardware utilization.