融合多尺度特征的目标检测模型 
下载： 846次
Object Detection Model Based on Multi-Scale Feature Integration
辽宁工程技术大学软件学院, 辽宁 葫芦岛 125105
图 & 表
图 1. RF-YOLOv2检测流程图
Fig. 1. Flowchart of RF-YOLOv2 detection
下载图片 查看原文
图 2. 目标函数变化曲线
Fig. 2. Object function change curve
下载图片 查看原文
图 3. 残差块结构
Fig. 3. Residual block structure
下载图片 查看原文
图 4. 特征金字塔网络
Fig. 4. Feature pyramid network
下载图片 查看原文
图 5. RF-YOLOv2流程图
Fig. 5. Flowchart of RF-YOLOv2
下载图片 查看原文
图 6. 各类别在KITTI数据集上出现的数量
Fig. 6. Number of categories appearing on KITTI data set
下载图片 查看原文
图 7. 两种模型的损失图
Fig. 7. Loss graph for two models
下载图片 查看原文
图 8. 两种模型的Precision-Recall曲线图。(a)(c)(e) YOLOv2模型;(b)(d)(f) RF-YOLOv2模型
Fig. 8. Precision-Recall curves of two models. (a)(c)(e) YOLOv2 model;(b)(d)(f) RF-YOLOv2 model
下载图片 查看原文
图 9. 检测结果图。(a)(c)(e)(g)(i) YOLOv2模型检测结果;(b)(d)(f)(h)(j) RF-YOLOv2模型检测结果
Fig. 9. Detection results. (a)(c)(e)(g)(i) Detection results of YOLOv2 model; (b)(d)(f)(h)(j) detection results of RF-YOLOv2 model
下载图片 查看原文
表 1RF-YOLOv2 网络结构
Table1. RF-YOLOv2 network structure
| Layerblock | Type | Numberof filters | Size /stride | Output |
|---|
| Convolutional | 32 | 3×3 | 416×416 | | Maxpool | | 2×2/2 | 208×208 | | Convolutional | 64 | 3×3 | 208×208 | | 1× | Convolutional | 32 | 1×1 | | | Convolutional | 64 | 3×3 | | | Residual | | | 208×208 | | Maxpool | | 2×2/2 | 104×104 | | Convolutional | 128 | 3×3 | 104×104 | | 2× | Convolutional | 64 | 1×1 | | | Convolutional | 128 | 3×3 | | | Residual | | | 104×104 | | Maxpool | | 2×2/2 | 52×52 | | Convolutional | 256 | 3×3 | 52×52 | | 4× | Convolutional | 128 | 1×1 | | | Convolutional | 256 | 3×3 | | | Residual | | | 52×52 | | Maxpool | | 2×2/2 | 26×26 | | Convolutional | 512 | 3×3 | 26×26 | | 4× | Convolutional | 256 | 1×1 | | | Convolutional | 512 | 3×3 | | | Residual | | | 26×26 | | Maxpool | | 2×2/2 | 13×13 | | Convolutional | 1024 | 3×3 | 13×13 | | 4× | Convolutional | 512 | 1×1 | | | Convolutional | 1024 | 3×3 | | | Residual | | | 13×13 | | Avgpool | | Global | 3 | | Softmax | | | |
|
查看原文
表 2精确率和检测速度对比
Table2. Comparison of accuracy and detection speed
| Model | Accuracyofcar /% | Accuracy ofpedestrian /% | Accuracy ofcyclist /% | Detectionspeed /(frame·s-1) |
|---|
| YOLOv2 | 68.56 | 44.26 | 55.95 | 46.4 | | RF-YOLOv2 | 87.88 | 52.91 | 74.05 | 30.3 | | YOLOv3 | 89.34 | 60.93 | 83.94 | 23.1 |
|
查看原文
表 3召回率和交并比的变化过程
Table3. Change process of recall rate and IOU
| Number oftraining | RF-YOLOv2 model | YOLOv2 model |
|---|
| Recallrate /% | IOU /% | Recallrate /% | IOU /% |
|---|
| 10000 | 50.36 | 43.29 | 48.18 | 43.42 | | 20000 | 55.45 | 46.34 | 53.11 | 45.98 | | 30000 | 61.47 | 50.65 | 55.83 | 47.79 | | 40000 | 64.92 | 52.56 | 54.13 | 46.72 | | 50000 | 65.87 | 53.63 | 57.98 | 49.04 |
|
查看原文
表 4car类别三种样本检测结果
Table4. Three sample detection results of car category
| Model | Accuracy of easy sample /% | Accuracy of moderate sample /% | Accuracy of hard sample /% |
|---|
| YOLOv2 | 70.56 | 57.32 | 50.44 | | Faster-rcnn | 87.90 | 79.11 | 70.19 | | RF-YOLOv2 | 91.01 | 81.26 | 72.41 |
|
查看原文
表 5pedestrian类别三种样本检测结果
Table5. Three sample detection results of pedestrian category
| Model | Accuracy of easy sample /% | Accuracy of moderate sample /% | Accuracy of hard sample /% |
|---|
| YOLOv2 | 59.97 | 49.05 | 44.91 | | Faster-rcnn | 78.35 | 65.91 | 61.19 | | RF-YOLOv2 | 64.35 | 57.02 | 53.94 |
|
查看原文
表 6cyclist类别三种样本检测结果
Table6. Three sample detection results of cyclist category
| Model | Accuracy of easy sample /% | Accuracy of moderate sample /% | Accuracy of hard sample /% |
|---|
| YOLOv2 | 56.47 | 56.68 | 53.02 | | Faster-rcnn | 71.41 | 62.81 | 55.44 | | RF-YOLOv2 | 79.76 | 74.68 | 72.41 |
|
查看原文
刘万军, 王凤, 曲海成. 融合多尺度特征的目标检测模型[J]. 激光与光电子学进展, 2019, 56(23): 231007. Wanjun Liu, Feng Wang, Haicheng Qu. Object Detection Model Based on Multi-Scale Feature Integration[J]. Laser & Optoelectronics Progress, 2019, 56(23): 231007.
PDF全文
