基于改进SSD的交通大场景多目标检测 下载: 1646次
Multi-Objective Detection of Traffic Scenes Based on Improved SSD
1 中国人民解放军陆军工程大学野战工程学院, 江苏 南京 210007
2 南部战区陆军第二工程科研设计所, 云南 昆明 650222
图 & 表
图 1. 改进后检测算法整体框架
Fig. 1. Improved detection algorithm overall framework
下载图片 查看原文
图 2. CNN模型提取各级特征卷积核示例
Fig. 2. Example of CNN model extracting feature convolution kernels at various levels
下载图片 查看原文
图 3. 二维Gabor滤波器卷积核
Fig. 3. Two-dimensional Gabor filter convolution kernel
下载图片 查看原文
图 4. 三维Gabor滤波器卷积核
Fig. 4. Three-dimensional Gabor filter convolution kernel
下载图片 查看原文
图 5. 最优Gabor卷积核组的训练流程
Fig. 5. Training process for optimal Gabor convolution kernel group
下载图片 查看原文
图 6. 基于时间感知特征映射的移动视频目标检测框架
Fig. 6. Mobile video target detection framework based on time-aware feature mapping
下载图片 查看原文
图 7. 模型在处理视频输入和输出示意图
Fig. 7. Model processing video input and output schematics
下载图片 查看原文
图 8. M4模型检测结果示例
Fig. 8. Example of M4 model detection results
下载图片 查看原文
表 1各模型识别和检测效果比较
Table1. Comparison of model identification and detection effects
Model | Dataset | AP /% | mAP /% | Pf /% | Pm /% | Pd /% | Pe /% |
---|
| | | Person | Car | Cyclist |
---|
M0 | KITTI | 73.36 | 71.53 | 65.32 | 70.07 | 20.21 | 19.34 | 41.32 | 19.13 | | WD | 71.59 | 69.63 | 62.75 | 67.99 | 19.25 | 21.38 | 38.83 | 20.54 | M1 | KITTI | 87.53 | 82.16 | 78.28 | 82.66 | 16.48 | 17.91 | 57.38 | 8.23 | | WD | 85.64 | 80.59 | 74.34 | 80.19 | 18.95 | 19.28 | 51.42 | 10.35 | M2 | KITTI | 77.18 | 72.35 | 68.69 | 72.74 | 12.31 | 13.29 | 57.84 | 16.56 | | WD | 73.52 | 70.45 | 64.83 | 69.61 | 15.17 | 14.49 | 52.45 | 17.89 | M3 | KITTI | 88.42 | 81.73 | 74.38 | 81.51 | 9.53 | 11.69 | 64.25 | 14.53 | | WD | 74.92 | 72.34 | 65.63 | 70.96 | 16.24 | 15.19 | 51.16 | 17.41 | M4 | KITTI | 92.42 | 92.23 | 90.85 | 91.83 | 5.19 | 7.13 | 81.47 | 6.21 | | WD | 88.46 | 87.38 | 83.24 | 86.36 | 8.26 | 11.27 | 71.05 | 9.42 |
|
查看原文
表 2不同算法检测和识别效果比较
Table2. Comparison of detection and recognition with different algorithms
Method | Dataset | AP /% | mAP /% | Pd /% | FPS /(frame·s-1) |
---|
| | | Person | Car | Cyclist |
---|
Faster R-CNN | KITTI | 83.26 | 74.13 | 75.42 | 77.61 | 45.22 | 13.15 | | WD | 81.49 | 71.33 | 68.65 | 73.82 | 36.63 | 11.64 | DSOD300 | KITTI | 77.43 | 72.26 | 68.38 | 72.69 | 58.68 | 58.23 | | WD | 70.73 | 69.39 | 67.04 | 69.05 | 52.32 | 50.35 | DSSD513 | KITTI | 75.46 | 69.53 | 68.34 | 71.11 | 59.42 | 46.34 | | WD | 72.19 | 68.83 | 66.45 | 69.16 | 49.79 | 39.38 | YOLOv2 544 | KITTI | 79.43 | 71.25 | 67.32 | 72.66 | 60.82 | 56.74 | | WD | 73.29 | 69.63 | 68.85 | 70.59 | 54.86 | 49.28 | M4 | KITTI | 92.42 | 92.23 | 90.85 | 91.83 | 81.47 | 31.86 | | WD | 88.46 | 87.38 | 83.24 | 86.36 | 71.05 | 19.83 |
|
查看原文
华夏, 王新晴, 王东, 马昭烨, 邵发明. 基于改进SSD的交通大场景多目标检测[J]. 光学学报, 2018, 38(12): 1215003. Xia Hua, Xinqing Wang, Dong Wang, Zhaoye Ma, Faming Shao. Multi-Objective Detection of Traffic Scenes Based on Improved SSD[J]. Acta Optica Sinica, 2018, 38(12): 1215003.