Small Target Tracking and Failure Discrimination Method Based on Improved KCF Under Object Detection and Instruction Guidance Framework

Tracking small targets in complex ground scenes is challenging due to insufficient feature representation and low contrast with the background. Traditional Kernelized Correlation Filter (KCF) methods struggle with small target representation due to fixed-resolution feature maps. Furthermore, failure detection based on a single tracking indicator and hard thresholding often leads to misjudgment, resulting in tracking failures. This paper proposes a tracking method that integrates dynamic feature map adjustment with online failure detection. By constructing a detection-tracking collaborative framework, feature map sizes are adjusted based on feedback from target detection, enhancing tracking stability. A polynomial interpolation-based target positioning method is designed to improve tracking accuracy. Additionally, an online Linear Discriminant Analysis (LDA) classifier is introduced to construct a robust failure determination model incorporating features such as response map kurtosis, enabling probabilistic evaluation of tracking states. Experimental results demonstrate that the proposed method significantly enhances tracking stability and accuracy compared to the conventional KCF algorithm. It also possesses real-time failure detection capabilities for occluded and camouflaged targets, providing a novel technological approach for continuous tracking of small targets in applications such as UAV reconnaissance and intelligent security systems.