Anti-Occlusion Target Tracking Algorithm Based on Fusion of Deep Features and Handcrafted Features

Abstract

This paper proposes an anti-occlusion target tracking algorithm that integrates deep convolutional features with handcrafted features and adds Average Peak Correlation Energy(APCE). The performance of traditional handcrafted features, such as Histogram of Oriented Gradient(HOG) feature, is unsatisfactory in complex environments. This paper uses deep convolutional features with HOG feature and Color Naming(CN) feature, Fully consider the characteristics of deep convolutional feature with strong representation ability and the characteristics of handcrafted feature extraction is simple. For the target occlusion problem, the APCE is introduced to evaluate the reliability of the tracking target. Once the target is occluded, the filter stops updating the target model and searches the target again. The results tested on OTB-100 video sequence set demonstrates that the improved algorithm has better performance accuracy and success rate than Kernel Correlation Filter(KCF) algorithm in occlusion and motion blur scene.