Joint information fusion and multi-scale network model for pedestrian detection
- PDF / 2,961,458 Bytes
- 10 Pages / 595.276 x 790.866 pts Page_size
- 18 Downloads / 189 Views
ORIGINAL ARTICLE
Joint information fusion and multi-scale network model for pedestrian detection Hexiang Zhang1 · Ziyu Hu2
· Ruoxin Hao1
Accepted: 12 October 2020 © Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract The existing pedestrian detection suffers the low accuracy when the environment changes dramatically. In order to solve the problem, a pedestrian detection model combining information fusion and multi-scale detection is proposed. The model is composed of a retinex algorithm and an improved YOLOv3 algorithm. Retinex algorithm is selected as the preprocessing algorithm to improve the brightness and contrast of pedestrians. The model improves the YOLOv3 algorithm by adding multiple scale detections. The K-means is used to determine the number of optimal anchors and the aspect ratio. By testing on the standard data set, the mean average precision (mAP) of the joint detection model increases from the original 80.69– 91.07%, and the recall increases from 65.22 to 87.48%. The comparative experiments show that the improved model performs good robustness and generalization ability on the problem of low pedestrian detection accuracy in complex environments. Keywords YOLOv3 · Pedestrian detection · Information fusion · Multi-scale network
1 Introduction Pedestrian detection [1] is a core technology in some applications, such as application in robot vision [2], intelligent monitoring systems, and driverless assistance systems [3], etc. Pedestrian detection technology can determine whether there is a pedestrian in the input image and the approximate location of the pedestrian. The accuracy and real-time performance of pedestrian detection is essential for the evaluation of the superiority of the detection system. Pedestrian detection belongs to the existing category of object detection, and the development of pedestrian detection can be divided into the following three periods. (1) The first period is the traditional object detection algorithm, which mainly uses artificial design features for
B
Ziyu Hu [email protected] Hexiang Zhang [email protected]
1
Institute of Electrical Engineering, Yanshan University, Qinhuangdao 066004, China
2
Engineering Research Center of the Ministry of Education for Intelligent Control System and Intelligent Equipment, Yanshan University, Qinhuangdao, China
global feature pedestrian detection. Features are generally extracted by artificial design methods, such as Harr, LBP, HOG [4,5] features, Shapelet and Edgelte features, etc. Support vector machine classifier is mainly used in the classification stage. According to the traditional detection algorithm, Navneet Dalal and Bill Triggs proposed the HOG + SVM [6,7] network structure to improve detection accuracy and robustness, which makes traditional object detection possible for pedestrian detection. Since it is difficult for artificially designed features to express the features of pedestrians in complex environments, the model detection accuracy is low. The algorithm itself has program redundancy and cannot be d
Data Loading...
