视觉传感器:2D感知算法
新机器视觉
共 6763字,需浏览 14分钟
· 2021-11-30
点击下方卡片,关注“新机器视觉”公众号
视觉/图像重磅干货,第一时间送达
1 前言
2 物体检测
2.1 两阶段检测
![](https://filescdn.proginn.com/52814fa94bb0aa5731088c739a321ac5/d68b5c99ba7a8df73994b6e5ac801a6f.webp)
![](https://filescdn.proginn.com/1bdc0e69900efe6c03bbc58acdefcdf2/d0794d0f0a2177bc1039fb1f7bb07d97.webp)
![](https://filescdn.proginn.com/4fb4368e81056ad792afd087a4b88c67/34de1f07f9eac5c0309d05260ee382f1.webp)
![](https://filescdn.proginn.com/398a9867755e1cfd3aef0b17caaf430c/e13c438e0506c0e8e3c4097ab5e16596.webp)
2.2 单阶段检测
![](https://filescdn.proginn.com/4c0c4d84b7dbb55f7beb2b9f1266f8ad/0801358acec8ed07608154081dc7adbe.webp)
![](https://filescdn.proginn.com/4f2ccdee2640dc97abf999b5d3681f48/9327e7fb7d06f86ca0c58e75c4be326a.webp)
![](https://filescdn.proginn.com/891dee36687568d1e6d1457858903c36/6e5a0a469d87162e1a0c6fbf8822dca8.webp)
2.3 无Anchor检测
![](https://filescdn.proginn.com/c0e40d6178572dd01da87890eca52359/7a015fc1567af0e29e1ccea0e23ad42a.webp)
![](https://filescdn.proginn.com/e645207d71b916b8fab16deb81ec0ba4/e7591ef365aa7acdb84790d5275f6431.webp)
![](https://filescdn.proginn.com/954af0a637d60f1e8e957d68acfabfa7/027424706ce5da4e31d312f2f7cbbc4c.webp)
![](https://filescdn.proginn.com/2de8983e78daaaf0d054fafa5e41c48f/0f615fb87401fd60567d2af54704f673.webp)
2.4 性能对比
![](https://filescdn.proginn.com/172913429399ebc853fb58f35ad69b0d/b75a334eaba762bf1dd00c27cf4b4f20.webp)
3 物体跟踪
由物体检测器在单帧图像上得到物体框输出。 提取每个检测物体的特征,通常包括视觉特征和运动特征。 根据特征计算来自相邻帧的物体检测之间的相似度,以判断其来自同一个目标的概率。 将相邻帧的物体检测进行匹配,给来自同一个目标的物体分配相同的ID。
![](https://filescdn.proginn.com/f71521ae0b4135dffbff6a0856de5204/5522bc3c228a0ca16462c599b0bbe095.webp)
![](https://filescdn.proginn.com/5455cd030f2726ff2d19148219f742b8/c8262e9063ce06d14705b47807719e18.webp)
![](https://filescdn.proginn.com/7b1219e9016777126818f7813e3c11d8/1cc1abe59d23582a309865e74c1ba535.webp)
![](https://filescdn.proginn.com/1505fa1072d88f82a00b59aa89c71a20/c154efb104c073fee9ada5648295b32d.webp)
4 语义分割
![](https://filescdn.proginn.com/3a718d16cddf6e7b0e4661ed97beab52/be43e04f9010d6847b134e0b35f9a527.webp)
![](https://filescdn.proginn.com/f7268cbeef8d92330504fbd04045a7ac/2af29fb6606433eddbaca004a0e690fa.webp)
![](https://filescdn.proginn.com/b5a5832dabfc0ab2e08fffde359656ca/4ab3e03bc771470f6679fac257db42c6.webp)
![](https://filescdn.proginn.com/9950502c3412c22922274f643e4c99b7/f5da8a9fcfb83373df9041bcf7b3e3c0.webp)
![](https://filescdn.proginn.com/1823ffb26726390c5f4af88f3fbb07ae/c9a885135e690128df922fef4fed1218.webp)
参考文献:
[1] Girshick et al., Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation, 2014.
[2] Girshick, Fast R-CNN, 2015.
[3] Ren et al., Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks, 2016.
[4] Lin et al., Feature Pyramid Networks for Object Detection, 2017.
[5] Liu et al., SSD: Single Shot MultiBox Detector, 2015.
[6] Lin et al., Focal Loss for Dense Object Detection, 2017.
—版权声明—
仅用于学术分享,版权属于原作者。
若有侵权,请联系微信号:yiyang-sy 删除或修改!
评论