| 摘 要: 为了在密集人群场景下提高检测精度,提出了一种基于YOLOv8改进的检测算法。首先,采用SENetV2作为主干网络,通过调整卷积网络的终端通道关系,以提升模型性能;其次,引入跨尺度特征融合模块CCFM(Cross-Scale Feature Fusion Module),将不同尺度的特征进行有效融合,增强模型对尺度变化的适应能力及对小尺度对象的检测能力。使用 Dynamic Head检测头,增加自注意力机制,以提高检测精度。此外,应用Slide Loss损失函数,以解决简单样本与困难样本之间的不平衡问题,从而进一步提升性能。实验结果显示,相较于YOLOv8n,改进后的算法在Crowd Human数据集上的mAP50和mAP95分别提升了3.9个百分点和3.2个百分点,同时,在 Wider Person数据集上,也提升了2.0个百分点和3.3个百分点。改进后的算法模型参数量减小了21.1%,实现了高精度与轻量化的良好平衡。 |
| 关键词: 密集人群 目标检测 YOLOv8 轻量化 |
|
中图分类号:
文献标识码: A
|
| 基金项目: 中共甘肃省委宣传部2023年农村电影公益放映第三方监测服务项目(2023zfcg00315) |
|
| Improved Dense Crowd Detection Algorithm Based on YOLOv8 |
|
CHEN Lintao, LIU Qiang
|
(College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China)
2074797942@qq.com; liuq@gasu.edu.cn
|
| Abstract: To improve the detection accuracy in dense crowd scenarios, an improved detection algorithm based on YOLOv8 is proposed. Firstly, SENetV2 is adopted as the backbone network to enhance model performance by adjusting the channel relationships at the terminal of the convolutional network. Secondly, the Cross-Scale Feature Fusion Module (CCFM) is introduced to effectively integrate features at different scales, strengthening the model’s adaptability to scale variations and its ability to detect smal-l scale objects. The Dynamic Head detection is employed, incorporating a sel-f attention mechanism to improve detection accuracy. Additionally, the Slide Loss function is applied to address the imbalance between simple and hard samples, further enhancing performance. Experimental results show that compared to YOLOv8n, the improved algorithm achieves increases of 3.9 percentage points and 3.2 percentage points in mAP 50 and mAP 95, respectively, on the Crowd Human dataset, while also improving by 2.0 percentage points and 3.3 percentage points on the Wider Person dataset. The parameter count of the improved model is reduced by 21.1% , achieving an excellent balance between high accuracy and lightweight design. |
| Keywords: dense crowds object detection YOLOv8 lightweight |
