大视场域的目标检测与识别算法综述 
下载： 2026次
Review on Object Detection and Recognition in Large Field of View
1 中国科学院上海微系统与信息技术研究所微系统技术重点实验室, 上海 201800
2 中国科学院大学, 北京 100049
图 & 表
图 1. 目标的非一致畸变
Fig. 1. Inconsistent distortion of object
下载图片 查看原文
图 2. 目标在LFOV图像中的残缺、模糊情况
Fig. 2. Incomplete and blurring object in LFOV images
下载图片 查看原文
图 3. 小目标在LFOV图像中的失真情况
Fig. 3. Distortion of small object in LFOV images
下载图片 查看原文
图 4. 目标在不同LFOV图像中的非对称性情况
Fig. 4. Asymmetry of object in different LFOV images
下载图片 查看原文
图 5. 大视场域目标检测与识别算法的分类
Fig. 5. Classification of object detection and recognition in large field of view
下载图片 查看原文
图 6. 基于畸变校正的目标检测与识别流程图
Fig. 6. Flow chart of object detection and recognition based on distortion correction
下载图片 查看原文
图 7. 大视场域的目标检测与识别算法的主要发展脉络
Fig. 7. Road map of object detection and recognition in large field of view
下载图片 查看原文
表 1各算法的性能比较
Table1. Comparison of performance of different algorithms
| Paper | Object | Pre-process | Feature extraction | Evaluation | Year |
|---|
| Jeong et al.[22] | Vehicle | Undistortion using FOV | HOG | | 2016 | | Silbersteinet al.[24] | Pedestrian | Camera calibration | AFS | 10.3%(avgMRF) | 2014 | | Levi andSilberstein[25] | Pedestrian | Camera calibration | AFS-Multi-Cue | 3.5%(avgMRF) | 2015 | | Bertozziet al.[26] | Pedestrian | Image un-warping | Soft-Cascadeand ACF | 0.35(FPPI is 10-1),0.59 (FPPI is 10-2),0.75 (FPPI is 10-3) | 2015 | | Suhr et al.[28] | Pedestrian | Image warping | HOG andTER-based classifier | 97.3%(TWA) | 2017 |
|
查看原文
表 2各算法的性能比较
Table2. Comparison of performance of different algorithms
| Paper | Object | Pre-process | Feature extraction | Evaluation | Year |
|---|
| Martinezet al.[31] | Humanbeings | Transformation | Viola-Jonesclassifier | Time reducedfrom600-700 ms to10-15 ms | 2010 | | Dinget al.[32] | Motionobject | | Based ontexture andcolor feature | 5× faster | 2016 |
|
查看原文
表 3各算法的性能比较
Table3. Comparison of performance of different algorithms
| Paper | Object | Pre-process | Feature extraction | Evaluation | Date |
|---|
| Yoshimiet al.[36] | Pedestrian | HPM | Faster R-CNN | LAMR: 24.5%;E=48.65% | 2017 | | Cai et al.[37] | VOC Pascal | Cylindrical unwarpingCorrection | YOLO | Detection rate: 30.89 frame·s-1Accuracy rate: 72% | 2018 | | Xu et al.[38] | Face | Spherical projection | LCNN | 97.3%(TWA) | 2018 | | Deng et al.[40] | Fire | Spherical projection | CNN | | 2017 |
|
查看原文
表 4各类畸变校正算法的性能比较
Table4. Comparison of performance of different algorithms
| Method | Advantage | Disadvantage |
|---|
| Camera calibration | Higher precision | Complexity | | No restrictions on camera type | Need to know a certain size of calibration object | | Non-metric calibration | Simple | Worse stability | | Few parameters | Not suitable for high-precision system |
|
查看原文
表 5各算法的性能比较
Table5. Comparison of performance of different algorithms
| Paper | Object | Method | Evaluation | Year |
|---|
| Deng et al.[42] | 20 class | Faster R-CNN | mAP: 68.7% | 2017 | | Coors et al.[49] | Flying cars | Encode the invariance of geometrictransformation directly into CNN | mAP: 50.18% | 2018 | | Herceg et al.[62] | Motion object | Corner detection, optical flow | Accuracy: 97.19% | 2011 | | Wu et al.[65] | Motion object | Moving blob method; PTZ shots image | Accuracy: 92% | 2017 | | Zhang et al.[67] | Salient object | Co-saliency detection algorithm | Precision: 0.82;recall: 0.75; F1: 0.81 | 2017 | | Cinaroglu et al.[69] | People | Omnidirectional sliding windowModified HOG+SVM | SVM scores: 2.94 | 2014 | | Wang et al.[70] | People | Template-based | MOTA: 0.85 | 2017 |
|
查看原文
李唐薇, 童官军, 李宝清, 卢晓洋. 大视场域的目标检测与识别算法综述[J]. 激光与光电子学进展, 2020, 57(12): 120002. Tangwei Li, Guanjun Tong, Baoqing Li, Xiaoyang Lu. Review on Object Detection and Recognition in Large Field of View[J]. Laser & Optoelectronics Progress, 2020, 57(12): 120002.
PDF全文
