基于點(diǎn)云補(bǔ)全的三維目標(biāo)檢測(cè)
2023年電子技術(shù)應(yīng)用第8期
陳輝,,王帥杰,,蔡晗
(桂林電子科技大學(xué) 信息與通信學(xué)院,, 廣西 桂林 541004)
摘要: LiDAR技術(shù)的發(fā)展為自動(dòng)駕駛提供了豐富的3D數(shù)據(jù),。然而,,由于遮擋和某些反射材料的原因引起信號(hào)丟失,,LiDAR點(diǎn)云實(shí)際上是不完整的2.5D數(shù)據(jù),,這對(duì) 3D 感知提出了根本性挑戰(zhàn)。針對(duì)這一問題,,提出對(duì)原始數(shù)據(jù)進(jìn)行三維補(bǔ)全的方法,。根據(jù)大多數(shù)物體形狀對(duì)稱且重復(fù)率高的特點(diǎn),通過學(xué)習(xí)先驗(yàn)對(duì)象形狀的方法估計(jì)點(diǎn)云中遮擋部分的完整形狀,。該方法首先識(shí)別被遮擋和信號(hào)缺失影響的區(qū)域,,在這些區(qū)域中預(yù)測(cè)區(qū)域所包含對(duì)象形狀的占用概率。針對(duì)物體間遮擋的情況,,通過形狀的占用概率和共享同類形狀形態(tài)進(jìn)行三維補(bǔ)全,。對(duì)自身遮擋的物體,通過自身鏡像進(jìn)行恢復(fù),。最后通過點(diǎn)云目標(biāo)檢測(cè)網(wǎng)絡(luò)進(jìn)行學(xué)習(xí)。結(jié)果表明,,通過該方法能有效地提高生成點(diǎn)云3D邊框的mAP(mean Average Precision),。
中圖分類號(hào):TP389.1
文獻(xiàn)標(biāo)志碼:A
DOI: 10.16157/j.issn.0258-7998.223624
中文引用格式: 陳輝,王帥杰,,蔡晗. 基于點(diǎn)云補(bǔ)全的三維目標(biāo)檢測(cè)[J]. 電子技術(shù)應(yīng)用,,2023,49(8):1-6.
英文引用格式: Chen Hui,,Wang Shuaijie,,Cai Han. 3D object detection based on point cloud completion[J]. Application of Electronic Technique,2023,,49(8):1-6.
文獻(xiàn)標(biāo)志碼:A
DOI: 10.16157/j.issn.0258-7998.223624
中文引用格式: 陳輝,王帥杰,,蔡晗. 基于點(diǎn)云補(bǔ)全的三維目標(biāo)檢測(cè)[J]. 電子技術(shù)應(yīng)用,,2023,49(8):1-6.
英文引用格式: Chen Hui,,Wang Shuaijie,,Cai Han. 3D object detection based on point cloud completion[J]. Application of Electronic Technique,2023,,49(8):1-6.
3D object detection based on point cloud completion
Chen Hui,,Wang Shuaijie,Cai Han
(School of lnformation and Communication,, Guilin University of Electronic Technology,, Guilin 541004, China)
Abstract: The development of LiDAR technology provides abundant 3D data for autonomous driving. However, LIDAR point cloud is actually incomplete 2.5D data due to signal loss caused by occlusion and some reflective materials, which poses a fundamental challenge to 3D perception. To solve this problem, this paper proposes a method for 3D completion of the original data. According to the symmetric shape and high repetition rate of most objects, the complete shape of the occluded part in the point cloud is estimated by learning the prior object shape. The method first identifies regions affected by occlusions and signal loss, and in these regions, predicts the occupancy probability of the shapes of objects contained in the regions. For the case of occlusion between objects, 3D completion is performed through the occupancy probability of the shape and the morphologies that share the same shape. The objects occluded by themselves are restored by mirroring themselves. Finally, it is learned through the point cloud target detection network. The results show that this method can effectively improve the mAP for generating point cloud 3D borders.
Key words : LiDAR,;point cloud,;3D completion;target detection
0 引言
3D目標(biāo)檢測(cè)作為自動(dòng)駕駛感知系統(tǒng)的核心基礎(chǔ)之一,,可以廣泛應(yīng)用于路徑規(guī)劃,、運(yùn)動(dòng)預(yù)測(cè)、碰撞避免等功能,。通常,,帶有相應(yīng)3D激光雷傳感器的汽車已經(jīng)成自動(dòng)駕駛領(lǐng)域的標(biāo)準(zhǔn)配置,由此能夠提供準(zhǔn)確的深度信息,,點(diǎn)云數(shù)據(jù)的處理也越來越普遍,、越來越重要,。盡管已有很多進(jìn)展,但由于點(diǎn)云本質(zhì)上的高度稀疏性和不規(guī)則的特性,,使得傳統(tǒng)的卷積神經(jīng)網(wǎng)絡(luò)無法對(duì)點(diǎn)云數(shù)據(jù)進(jìn)行準(zhǔn)確的學(xué)習(xí),,而且由于相機(jī)視圖和激光雷達(dá)鳥瞰視圖之間的不對(duì)齊而導(dǎo)致的導(dǎo)致模態(tài)協(xié)同和遠(yuǎn)距離尺度變化等原因,三維點(diǎn)云的處理遠(yuǎn)比二維圖像要難得多,。因此,,在三維點(diǎn)云上的目標(biāo)檢測(cè)目前仍處于初級(jí)階段。
本文詳細(xì)內(nèi)容請(qǐng)下載:http://forexkbc.com/resource/share/2000005474
作者信息:
陳輝,,王帥杰,,蔡晗
(桂林電子科技大學(xué) 信息與通信學(xué)院, 廣西 桂林 541004)
此內(nèi)容為AET網(wǎng)站原創(chuàng),,未經(jīng)授權(quán)禁止轉(zhuǎn)載,。