A New Siamese Heterogeneous Convolutional Neural Networks Based on Attention Mechanism and Feature Pyramid

Area of Science:

• Computer Vision
• Deep Learning
• Object Tracking

Background:

• Deep fully convolutional neural networks (CNNs) face challenges in object tracking, including tracking drift and reduced speed due to convolution padding, receptive field, and network step size.
• Existing methods struggle to balance accuracy and efficiency in complex tracking scenarios.

Purpose of the Study:

• To propose a novel fully convolutional siamese network object tracking algorithm that enhances both accuracy and speed.
• To address tracking drift and computational inefficiency in deep CNN-based trackers.

Main Methods:

• The proposed tracker utilizes a new fully CNN for feature extraction, incorporating a channel attention mechanism to improve feature representation.
• A feature pyramid network (FPN) fuses high and low-level convolutional features for similarity learning.
• Heterogeneous convolution kernels are employed to reduce computational complexity (FLOPs) and parameters, improving tracking speed.

Main Results:

• Experimental validation on VOT-2017, VOT-2018, OTB-2013, and OTB-2015 datasets demonstrates superior performance compared to state-of-the-art trackers.
• The integration of attention mechanisms and FPN effectively improved the representation and fusion of features.
• Heterogeneous convolution kernels successfully reduced computational load without compromising accuracy.

Conclusions:

• The developed object tracking algorithm achieves a significant improvement in both accuracy and speed.
• The combination of attention mechanisms, FPN, and heterogeneous convolutions offers a robust solution for efficient and accurate object tracking.