【摘要】：物體檢測技術(shù)是協(xié)助機器人視覺系統(tǒng)感知、定位視野中存在的客觀物體的關(guān)鍵技術(shù),是機器人視覺系統(tǒng)完成場景理解等高層視覺任務(wù)的基礎(chǔ)。研究并提升機器人視覺系統(tǒng)中的物體檢測算法的性能對提升機器人的智能程度具有重要意義。本文將機器人視覺系統(tǒng)中的物體檢測問題根據(jù)目標(biāo)物體的泛化程度劃分為三個層次:剛性物體檢測、習(xí)得物體檢測和任意物體檢測。然后遵循從簡單到復(fù)雜、從具體到泛化的順序?qū)@三類檢測問題依次展開研究。針對剛性物體檢測,本文結(jié)合所在課題組開發(fā)的兩種機器人(火電廠用溫度探頭組裝機器人和中國象棋對弈機器人)對物體檢測的實際工程需求,提出了兩種剛性物體檢測算法。該兩種算法已在各自的機械臂手眼系統(tǒng)中順利運行。然后,本文對習(xí)得物體檢測進行探究,尤其關(guān)注卷積神經(jīng)網(wǎng)絡(luò)在習(xí)得物體檢測中的應(yīng)用。針對具有廣泛應(yīng)用前景的自然場景下的二代身份證檢測問題,本文提出了一種可用于端到端的習(xí)得物體檢測的卷積神經(jīng)網(wǎng)絡(luò)模型。在自然場景圖片中的二代身份證檢測任務(wù)中,該模型在本文收集的數(shù)據(jù)集上取得了 79.51%的平均覆蓋率。最后,本文對任意物體檢測問題展開研究并創(chuàng)新地提出了基于圖像分割結(jié)構(gòu)化特征的任意物體檢測算法。該算法在候選集容量為1000時在權(quán)威數(shù)據(jù)集PASCAL VOC2007上取得了 96.1%的檢測率。在候選集容量小于100時,檢測性能優(yōu)于四種性能突出的經(jīng)典算法。
[Abstract]:Object detection technology is the key technology to assist the robot vision system to perceive and locate the objective objects in the visual field. It is also the foundation of the robot vision system to complete the high-level vision tasks such as scene understanding. It is very important to study and improve the performance of object detection algorithm in robot vision system. In this paper, the object detection problem in robot vision system is divided into three levels according to the generalization of target object: rigid object detection, acquisition object detection and arbitrary object detection. Then follow the order from simplicity to complexity and from concrete to generalization to study these three detection problems in turn. Aiming at the rigid object detection, this paper combines the actual engineering requirements of two kinds of robots (temperature probe assembled robot and Chinese chess game robot) developed by our research group for object detection. Two kinds of rigid object detection algorithms are proposed. The two algorithms have been running smoothly in the hand-eye system of each manipulator. Then, this paper explores acquisition object detection, especially the application of convolution neural network in acquisition object detection. Aiming at the problem of second-generation ID card detection in natural scenes with wide application prospects, a convolution neural network model for end-to-end acquisition object detection is proposed in this paper. In the second generation ID card detection task in natural scene images, the model achieves an average coverage of 79.51% on the data set collected in this paper. Finally, this paper studies the problem of arbitrary object detection and proposes an algorithm of arbitrary object detection based on the structured feature of image segmentation. The algorithm achieves a detection rate of 96.1% on the authoritative dataset PASCAL VOC2007 when the candidate set capacity is 1000. When the capacity of candidate set is less than 100, the detection performance is better than four classical algorithms with outstanding performance.
【學(xué)位授予單位】：浙江大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2017
【分類號】：TP391.41;TP242

【參考文獻】

相關(guān)期刊論文 前1條

1 王殿君;;基于視覺的中國象棋棋子識別定位技術(shù)[J];清華大學(xué)學(xué)報(自然科學(xué)版);2013年08期

，

本文編號：2231596