Research on video target tracking algorithm based on deep neural network

Abstract

: In the field of computer vision applications, visual object tracking is a widely researched and hot-topic area, finding extensive practical applications in many key visual domains and demonstrating promising real-world performance. However, due to various factors such as lighting variations, scale changes, background clutter, low resolution, and other interferences, visual object tracking requires improvements on multiple fronts. In this paper, a video object tracking algorithm based on deep neural networks is proposed while ensuring real-time tracking. Addressing the limitation of traditional visual object tracking algorithms based on correlation filtering theory, which rely on shallow handcrafted features, this algorithm first leverages a deep neural network model to extract deep features of the target to be tracked. Given that different convolutional layers encode different information in their deep feature representations, these distinct layer features are subsequently fused to enhance representation capability. Furthermore, a kernel correlation-based approach is employed to boost the tracking speed of the visual object tracking algorithm. The experimental results demonstrate that the method proposed in this paper achieves a balance between target tracking accuracy and speed, enhancing the robustness of visual object tracking algorithms in complex and noisy backgrounds.