Improved gesture detection algorithm based on YOLOv3

In order to solve the problem of low accuracy and serious redundancy of traditional gesture detection algorithms in first-view and multi-gesture scenarios, this paper proposes an improved gesture detection algorithm based on the YOLOv3 network framework. First, use the K-means++ clustering algorithm...

Full description

Saved in:
Bibliographic Details
Published in2021 40th Chinese Control Conference (CCC) pp. 7068 - 7073
Main Authors Zhan, Jinfeng, Liu, Weidong, Yang, Weirong
Format Conference Proceeding
LanguageEnglish
Published Technical Committee on Control Theory, Chinese Association of Automation 26.07.2021
Subjects
Clustering algorithms
Detection algorithms
Egohands
Focal Loss
Improved NMS
means
Real-time systems
Redundancy
Training
YOLOv3
Online AccessGet full text

Cover

More Information
Summary:In order to solve the problem of low accuracy and serious redundancy of traditional gesture detection algorithms in first-view and multi-gesture scenarios, this paper proposes an improved gesture detection algorithm based on the YOLOv3 network framework. First, use the K-means++ clustering algorithm instead of K-means to re-cluster the scale of the training set to obtain a set of Priors that are more in line with the gesture distribution; then, the Focal Loss is introduced to improve the loss function to solve the problem of extremely imbalanced positive and negative samples in the detection process, and let the network pay more attention to samples that are difficult to train; finally, improve the non-maximum suppression algorithm, get a result closer to ground truth by weighting the candidate boxes, and remove redundant candidate boxes by judging the positional relationship between the candidate bounding boxes. Experiments in the Egohands data set show that the algorithm proposed in this paper has a greater improvement in accuracy and mAP compared to the original YOLOv3 algorithm.
ISSN:2161-2927
DOI:10.23919/CCC52363.2021.9550573