Open Access

ARTICLE

MEB-YOLO: An Efficient Vehicle Detection Method in Complex Traffic Road Scenes

by Yingkun Song¹, Shunhe Hong¹, Chentao Hu¹, Pingan He², Lingbing Tao¹, Zhixin Tie^1,3,*, Chengfu Ding⁴

1 School of Computer Science and Technology, Zhejiang Sci-Tech University, Hangzhou 310018, China
2 School of Science, Zhejiang Sci-Tech University, Hangzhou 310018, China
3 Keyi College, Zhejiang Sci-Tech University, Shaoxing 312369, China
4 Focused Photonics (Hangzhou) Inc., Hangzhou 310052, China

* Corresponding Author: Zhixin Tie. Email:

Computers, Materials & Continua 2023, 75(3), 5761-5784. https://doi.org/10.32604/cmc.2023.038910

Received 03 January 2023; Accepted 16 March 2023; Issue published 29 April 2023

Abstract

Rapid and precise vehicle recognition and classification are essential for intelligent transportation systems, and road target detection is one of the most difficult tasks in the field of computer vision. The challenge in real-time road target detection is the ability to properly pinpoint relatively small vehicles in complicated environments. However, because road targets are prone to complicated backgrounds and sparse features, it is challenging to detect and identify vehicle kinds fast and reliably. We suggest a new vehicle detection model called MEB-YOLO, which combines Mosaic and MixUp data augmentation, Efficient Channel Attention (ECA) attention mechanism, Bidirectional Feature Pyramid Network (BiFPN) with You Only Look Once (YOLO) model, to overcome this problem. Four sections make up this model: Input, Backbone, Neck, and Prediction. First, to improve the detection dataset and strengthen the network, MixUp and Mosaic data improvement are used during the picture processing step. Second, an attention mechanism is introduced to the backbone network, which is Cross Stage Partial Darknet (CSPDarknet), to reduce the influence of irrelevant features in images. Third, to achieve more sophisticated feature fusion without increasing computing cost, the BiFPN structure is utilized to build the Neck network of the model. The final prediction results are then obtained using Decoupled Head. Experiments demonstrate that the proposed model outperforms several already available detection methods and delivers good detection results on the University at Albany DEtection and TRACking (UA-DETRAC) public dataset. It also enables effective vehicle detection on real traffic monitoring data. As a result, this technique is efficient for detecting road targets.

---

Keywords

---