Research on subway pedestrian detection algorithms based on SSD model

• Published in: IET intelligent transport systems Vol. 14; no. 11; pp. 1491 - 1496
• Main Authors: Yang, Jie, He, Wen Yu, Zhang, Tian Lu, Zhang, Chun Lei, Zeng, Lu, Nan, Bing Fei
• Format: Journal Article
• Language: English
• Published: The Institution of Engineering and Technology 01.11.2020; Wiley
• Subjects: base network; boundary box; crowded metro stations; feature extraction; feature information; feature map; flow statistics; image classification; input image; in‐depth learning detection; learning (artificial intelligence); local feature extraction; network structure; object categories; object detection; pedestriain flow impact; pedestrian target sizes; pedestrians; simulation experiments; single shot multibox detector; smart city application; Special Issue: Intelligent Transportation Systems in Smart Cities for Sustainable Environments; SSD model; subway pedestrian detection algorithms; subway station environment; target recognition; traffic engineering computing; VGG16; VOC2007; VOC2007; flow statistics; feature map; boundary box; image classification; in-depth learning detection; subway station environment; SSD model; VGG16; object detection; local feature extraction; feature extraction; input image; learning (artificial intelligence); network structure; pedestrian target sizes; pedestrians; subway pedestrian detection algorithms; traffic engineering computing; feature information; simulation experiments; target recognition; base network; single shot multibox detector; pedestriain flow impact; crowded metro stations; smart city application; object categories
• ISSN: 1751-956X; 1751-9578; 1751-9578
• DOI: 10.1049/iet-its.2019.0806

Accurate target recognition and location is one of the key technologies in the field of smart city application. In order to solve the problem of large pedestriain flow impact in crowded metro stations, a method of in-depth learning detection based on SSD (single shot multibox detector) is proposed. The algorithm extracts the feature information of the input image, then returns the boundary box of the location on the feature map and classifies the object categories. Using the method of local feature extraction, the features of different positions, different aspect ratios and sizes are obtained, and VGG16 is used as the base network to optimise and improve the network structure. The results of simulation experiments on VOC2007 and data_sub show that the maximum value of mAP is 77% and the highest accuracy is 96.31%. Compared with other mainstream deep learning target detection methods, SSD has higher accuracy, better real-time and robustness. It can solve the problem of different pedestrian target sizes and better realise pedestrians in subway station environment. Detection provides decision-making basis for flow statistics.