Object Tracking for Grabing Robot Based on Multiple Feature

Abstract

In order to solve problems encountered in capturing moving objects in the field of intelligent assembly, a modified algorithm based on correlation filtering and HOG is proposed. Facing these problems such as camera shake, object occlusion, manipulator motion planning to crash barrier, unable to advance object trajectory prediction, the algorithm in the past may have trouble. The HOG feature can't handle occlusion problems, and a method of multi feature fusion is put forward. The introduction of CN features, the use of target appearance color feature and tracking depth features instead of original HOG feature can better distinguish whether it is the target. Through the fusion of the two features, we improve the accuracy in target tracking, and proved that the improved algorithm has advantage. At the same time, we use AlexNet depth network, and select the second layer of convolution as the feature of target tracking. Through a lot of experiments, it is verified that the depth feature has high accuracy and robustness.