Affiliations:
College of Information and Control Engineering, Shenyang Jianzhu University, Shenyang, China
Correspondence:
[*]
Corresponding author. Liu Jian, College of Information and Control Engineering, Shenyang Jianzhu University,
Shenyang 110168, China. E-mail: xxliujian@sjzu.edu.cn.
Abstract: One difficulty that remains in image processing is the accurate location of key points in depth images. This paper presents an intelligent location method for identifying key points in depth images based on deep convolutional neural networks. This study used Kinect to process images, calculating the differences in depth as well as the directional gradient in subject depth images. The entirety of each depth image was traversed through a sliding window to identify the feature vector. Principal component analysis was used to reduce image dimensions. The random forest technique was used to select characteristics of strong classification as well as to actualize training and testing. A depth convolutional neural network was used to detect key points in images of pedestrians. During the study, an experimental test was conducted in a general environment under various conditions, including occlusion and low light. Even under these suboptimal conditions, the detection rate of the proposed method was 87.72%. Furthermore, this method was compared with the GEBCF and FCF algorithms, and proved to increase the detection rate by 0.92% and 0.68%, respectively. Using the depth convolutional neural network in the pedestrian key point positioning experiment, the average error obtained when comparing the predicted point coordinates to the sample mark coordinates was 2.102 pixels. These experimental results show that this method has good accuracy and robustness for the key point location problem of pedestrians in depth images.
Keywords: Pedestrian detection, depth image, key point location, deep learning
