DSiam-CnK: A CBAM- and KCF-Enabled Deep Siamese Region Proposal Network for Human Tracking in Dynamic and Occluded Scenes

Xiangpeng Liu ¹, Jianjiao Han ¹, Yulin Peng ¹

⁰College of Information, Mechanical & Electrical Engineering, Shanghai Normal University, 100 Haisi Road, Shanghai 201418, China.

Sensors (basel, Switzerland) +

|

January 8, 2025

View abstract on PubMed

Summary

This study introduces DSiam-CnK, a dynamic Siamese neural network that improves object tracking accuracy and robustness, especially during prolonged tracking and occlusions. It enhances template updates for better performance on challenging datasets.

Area Of Science

Computer Vision
Artificial Intelligence
Machine Learning

Background

Siamese neural networks in object tracking often rely heavily on initial frames, lacking dynamic template updates.
Prolonged tracking scenarios present challenges like target occlusion and frame exit for existing algorithms.

Purpose Of The Study

To enhance the SiamRPN algorithm for improved object tracking accuracy and robustness in prolonged scenarios.
To introduce dynamic template updating capabilities for adaptive tracking strategies.

Main Methods

Integration of the Convolutional Block Attention Module (CBAM) to boost spatial channel attention.
Incorporation of Kernelized Correlation Filters (KCFs) for superior feature template representation.
Development of DSiam-CnK, a Siamese network featuring dynamic template updating.

Main Results

On OTB2015, DSiam-CnK achieved a 92.1% success rate and 90.9% precision against SiamRPN.
VOT2018 results showed leading performance with 46.7% VOT-A, 135.3% VOT-R, and 26.4% VOT-EAO.
LaSOT dataset performance included 35.3% precision and 39% success rate.

Conclusions

The proposed DSiam-CnK method significantly enhances object tracking precision and robustness.
Dynamic template updating and attention mechanisms contribute to superior performance in challenging tracking conditions.

Keywords:

CBAM DSiam-CnK KCF SiamRPN object tracking template update