| 摘要: |
| 目的 解决道路交通环境中车辆目标较小、遮挡,以及目标不清晰、错检漏检等问题。 方法 提出一种用于评
估目标检测框与目标真实形状之间匹配程度的指标 ShapeIoU, 通过计算目标轮廓的相似性来衡量检测框与真实
形状的匹配程度,以提高目标检测的精度和准确性;引进动态蛇形卷积(DSConv)模块调节的卷积核模块,能够根
据目标形状和大小的变化自适应地调整卷积核形状,从而更有效地捕获目标的特征;为了提高多尺度检测能力,增
强检测精度以及模型的表达能力,改进中增加 P2 检测头。 结果 改进 YOLOv8 模型算法的 Precision、Recall、Map@
0. 5 以及 mAP@ 0. 5:0. 95 分别提高了 2. 6%、3. 2%、3. 2%和 3. 0%,达到了 87. 9%、84. 2%、89. 4%和 65. 4%。 结论
对于复杂的交通环境中车辆较小、遮挡,以及不清晰目标的检测有了明确提高,证实此改进方法有效解决了上述对
于车辆检测目标精度低等问题。 |
| 关键词: :YOLOv8 车辆检测 动态蛇形卷积 ShapeIoU P2 检测头 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Research on a Vehicle Detection Algorithm Based on Improved YOLOv8 |
|
CHEN Zhengzhi LIU Guowei
|
|
School of Electrical and Information Engineering Anhui University of Science and Technology Huainan 232001 Anhui
China
|
| Abstract: |
| Objective This study addresses the following challenges in vehicle detection under road traffic environments
small target sizes occlusions unclear targets and the consequent false positives and missed detections. Methods An
index named ShapeIoU was proposed to assess the matching degree between the target detection box and the real shape of
the target. By calculating the similarity of the target contour the matching degree between the detection box and the real
shape was measured to enhance the precision and accuracy of object detection. A convolutional kernel module adjusted by
the dynamic snake convolution DSConv module was introduced. This module could adaptively adjust the shape of the
convolutional kernel according to the changes in the shape and size of the target so as to capture the features of the target
more effectively. To improve the multi-scale detection ability enhance the detection accuracy and the expression ability of
the model a P2 detection head was added during the improvement. Results The Precision Recall mAP@ 0. 5 and
mAP@ 0. 5 0. 95 of the improved YOLOv8 model algorithm increased by 2. 6% 3. 2% 3. 2% and 3. 0% respectively
reaching 87. 9% 84. 2% 89. 4% and 65. 4%. Conclusion The detection of small occluded and unclear targets in
complex traffic environments has been significantly improved. It is verified that this improved method can effectively solve
problems such as low accuracy in vehicle target detection. |
| Key words: YOLOv8 vehicle detection dynamic snake convolution ShapeIoU P2 detection head |
