基于改进YOLOv3的实时人手检测算法

Real-Time Hand Detection Algorithm Based on Improved YOLOv3

针对目前人手检测方法在实时性和高精度方面难以均衡的情况,提出一种改进YOLOv3的实时人手检测算法。通过结合FPN和特征融合增加多尺度检测,并采用K-means算法得到预设的锚框值。改进算法在测试集上的平均检测精度达到57.61%,比原YOLOv3提升3.58%;检测速度每秒达到23.8帧,满足实时性需求。

Aiming at the situation that current hand detection methods are difficult to balance in terms of speed and accuracy,an improved real-time hand detection algorithm for YOLOv3 is proposed.Multi-scale detection is added by combining FPN and feature fusion,and the K-means algorithm is used to obtain preset anchor box values.The average detection accuracy of the improved algorithm on the test set reaches 57.61%,which is 3.58 percentage points higher than the original YOLOv3;the detection speed reaches 23.8 frames per second,which meets the real-time requirements.