李明明,雷菊陽*,趙從健



摘 ?要: 針對復(fù)雜道路場景的目標(biāo)檢測難以實(shí)現(xiàn)在移動設(shè)備上的實(shí)時目標(biāo)檢測問題,采用了MobileNet-SSD的目標(biāo)檢測框架,設(shè)計(jì)了一種用于視頻的多目標(biāo)檢測組合網(wǎng)絡(luò)框架LSTM-SSD。利用視頻連續(xù)幀的信息時序關(guān)聯(lián),有效的提高檢測的置信度,減少單一圖像檢測中存在的不穩(wěn)定問題。通過與VGG-SSD\MobileNet-SSD兩種檢測網(wǎng)絡(luò)模型的對比,實(shí)驗(yàn)表明,設(shè)計(jì)的檢測網(wǎng)絡(luò)模型在應(yīng)對多目標(biāo)、模糊、遮擋等干擾狀況下,均能獲得較好的檢測效果。該模型的設(shè)計(jì),可對無人駕駛實(shí)現(xiàn)實(shí)時目標(biāo)檢測提供依據(jù)和參考。
關(guān)鍵詞: 視頻多目標(biāo)檢測;SSD;時間維度特征;道路場景
中圖分類號: TP391.41 ? ?文獻(xiàn)標(biāo)識碼: A ? ?DOI:10.3969/j.issn.1003-6970.2019.12.031
本文著錄格式:李明明,雷菊陽,趙從健. 道路場景中基于視頻的多目標(biāo)檢測[J]. 軟件,2019,40(12):140145
Multi-target Detection Under Road Scenes Based on Video
LI Ming-ming, LEI Ju-yang*, ZHAO Cong-jian
(College of Mechanical and Automotive Engineering, Shanghai University of Engineering Science, Shanghai 201620, China)
【Abstract】: Aiming at the problem that it is difficult for mobile devices to realize real-time target detection of complex road scenes. based on MobileNet-SSD target detection framework, an LSTM-SSD combined model algorithm for multi-target detection of video is designed. The algorithm takes advantage of the temporal feature of the video to effectively improve the confidence of detection and reduce the instability problem in image detection. Compared with the two detection network models of VGG-SSD\MobileNet-SSD, the results show that the designed detection network model can obtain better detection results under multi-objective, fuzzy, occlusion and other interference conditions. The construction of the model can provide basis and reference for real-time target detection by driverless vehicles.
【Key words】: Video multi-target detection; SSD; Temporal feature; Road scenes
0 ?引言
無人駕駛是未來發(fā)展的重要方向,基于視覺的道路場景的目標(biāo)檢測是無人駕駛的主要研究課題[1]。在車輛行駛過程中,如何快速、準(zhǔn)確的檢測到車輛前方的行人、車輛、車道線、紅綠燈、提示牌等目標(biāo)物體,對無人駕駛系統(tǒng)提前制定駕駛方案具有重要的研究意義。
近幾年來,將深度學(xué)習(xí)應(yīng)用到目標(biāo)檢測方面取得了非常好的檢測效果。各專家學(xué)者提出了許多模型來解決視頻目標(biāo)檢測速度慢、精確度低的問題。Chen X[2]等提出了一種用于實(shí)時檢測的時間單發(fā)檢測器,開發(fā)的TSSD-OTA在檢測和跟蹤方面實(shí)現(xiàn)了快速和整體競爭性能。Liu[3]等具有時間感知特征映射的移動視頻目標(biāo)檢測,快速的單圖像目標(biāo)檢測模型與卷積長短期記憶(LSTM)層相結(jié)合,創(chuàng)造了混合的循環(huán)卷積體系結(jié)構(gòu)。……