Improved ECO Algorithm Based on Residual Neural Network

• Published in: Journal of physics. Conference series Vol. 1732; no. 1; pp. 12081 - 12090
• Main Authors: Qi, Huifang, Wang, Changmeng, Li, Bo
• Format: Journal Article
• Language: English
• Published: IOP Publishing 01.01.2021
• ISSN: 1742-6588; 1742-6596; 1742-6596
• DOI: 10.1088/1742-6596/1732/1/012081

Video target tracking is one of hot fields of computer vision, and its application is also very extensive. However, due to the complexity and variability of tracking environment, which brings some challenges to the research of target tracking. ECO(Efficient Convolution Operators) Algorithm is proposed based on convolutional neural network in three aspects. Firstly, residual neural network ResNet50 is adopted instead of convolutional neural network to extract the appearance features of target, and deeper residual neural network is applied to obtain more abundant target semantic information, so as to improve the tracking effect of tracking algorithm.Secondly, sample space classification strategy is improved. Different weights are assigned to shallow feature and deep feature, which make the deep feature play a more important role and improve the effect of target tracking. Finally, the method of scale estimation is improved so that better bounding boxes can be estimated with the scale changing. Experimental results show that the distance accuracy and success rate of the algorithm.