Open Access

ARTICLE

An Algorithm for Target Detection of Engineering Vehicles Based on Improved CenterNet

Pingping Yu¹, Hongda Wang¹, Xiaodong Zhao^1,*, Guangchen Ruan²

1 School of Information Science and Engineering, Hebei University of Science and Technology, Shijiazhuang, 050000, China
2 Indiana University at Bloomington, Indiana, 47405, USA

* Corresponding Author: Xiaodong Zhao. Email:

Computers, Materials & Continua 2022, 73(2), 4261-4276. https://doi.org/10.32604/cmc.2022.029239

Received 28 February 2022; Accepted 06 May 2022; Issue published 16 June 2022

Abstract

Aiming at the problems of low target image resolution, insufficient target feature extraction, low detection accuracy and poor real time in remote engineering vehicle detection, an improved CenterNet target detection model is proposed in this paper. Firstly, EfficientNet-B0 with Efficient Channel Attention (ECA) module is used as the basic network, which increases the quality and speed of feature extraction and reduces the number of model parameters. Then, the proposed Adaptive Fusion Bidirectional Feature Pyramid Network (AF-BiFPN) module is applied to fuse the features of different feature layers. Furthermore, the feature information of engineering vehicle targets is added by making full use of the high-level semantic and low-level fine-grained feature information of the target, which overcomes the problem that the original CenterNet network did not perform well in small target detection and improve the detection accuracy of the network. Finally, the tag coding strategy and bounding box regression method of CenterNet are optimized by introducing positioning quality loss. The accuracy of target prediction is increased by joint prediction of center position and target size. Experimental results show that the mean Average Precision (mAP) of the improved CenterNet model is 94.74% on the engineering vehicle dataset, and the detection rate is 29 FPS. Compared with the original CenterNet model based on ResNet-18, the detection accuracy of this model is improved by 16.29%, the detection speed is increased by 9 FPS, and the memory usage is reduced by 43 MB. Compared with YOLOv3 and YOLOv4, the mAP of this model is improved by 19.9% and 5.61% respectively. The proposed method can detect engineering vehicles more quickly and accurately in far distance. It has obvious advantages in target detection compared with traditional methods.

---

Keywords

---