Related Experiment Video
Updated: Sep 21, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
652
A New Deep Model for Detecting Multiple Moving Targets in Real Traffic Scenarios: Machine Vision-Based Vehicles
Xiaowei Xu1,2, Hao Xiong1,2, Liu Zhan1,2
1School of Automobile and Traffic Engineering, Wuhan University of Science and Technology, Wuhan 430081, China.
Sensors (Basel, Switzerland)
|May 28, 2022
Summary
This study introduces an improved YOLOv4 model for enhanced multiple target detection in complex traffic scenes. The enhanced model significantly boosts accuracy for small and occluded targets, achieving a higher mean average precision (mAP).
Area of Science:
- Computer Vision
- Machine Learning
- Deep Learning
Background:
- Detecting small and occluded targets in complex traffic scenes presents a significant challenge for current multiple target detection systems.
- Existing methods often struggle with accurately identifying objects that are partially hidden or appear small within the scene.
Purpose of the Study:
- To propose an improved YOLOv4 detection method that enhances the accuracy and efficiency of multiple target detection, particularly for small and occluded objects.
- To address the limitations of existing models in complex traffic environments.
Main Methods:
- An improved YOLOv4 model incorporating a four-scale detection structure by introducing a 4x down-sampling feature map into the neck network.
- Integration of the Convolutional Block Attention Module (CBAM) into the neck network to improve spatial and channel feature learning.
- Utilizing the soft non-maximum suppression (Soft-NMS) algorithm based on Distance Intersection over Union (DIoU) to optimize bounding box detection and reduce false negatives for occluded targets.
Main Results:
- The improved YOLOv4 model achieved a mean average accuracy (mAP) of 81.23% on the KITTI dataset, outperforming the original YOLOv4 by 3.18%.
- The model demonstrated a computation speed of 47.32 FPS, indicating efficient real-time processing capabilities.
- Comparative analysis showed superior detection accuracy and computation speed compared to existing popular detection models.
Conclusions:
- The proposed enhanced YOLOv4 model effectively improves multiple target detection accuracy in complex traffic scenes.
- The integration of architectural modifications, attention mechanisms, and advanced NMS algorithms leads to significant performance gains for small and occluded targets.
- The model offers a promising solution for real-world applications requiring robust and efficient object detection.
