新型工业化
新型工業化
신형공업화
New Industrialization Straregy
2013年
5期
102-112
,共11页
平躺人体检测%样本扩充%透视变换%支持向量机%梯度方向直方图
平躺人體檢測%樣本擴充%透視變換%支持嚮量機%梯度方嚮直方圖
평당인체검측%양본확충%투시변환%지지향량궤%제도방향직방도
Lying pose human detection%examples expanding%perspective transform%linear support vector machine%histogram of oriented gradient (HoG)
平躺人体检测具有很好的学术研究和应用价值。针对平躺人体表观姿态受透视变换的影响较大和样本收集难这两个问题展开研究：首先建立并公开了厦门大学平躺人体检测数据集，该数据集的人体包含各种姿态、视角、衣着颜色、复杂背景等情况；其次提出了一种基于透视变换的样本扩充方法：根据透视变换的摄像机模型将3x3的透视变换矩阵进行分解，使得该矩阵最终只与两个变化范围确定的旋转角相关，方便了采样算法的设计，有效地增加了训练样本的数目和样本表观的多样性，在厦门大学平躺人体数据集上的实验结果表明该方法提高了分类器的性能：当FPPW=10e-4时，漏检率降低了12.4%；当FPPW=10e-5时，漏检率降低了5.75%。可见本文提出的样本扩充方法简单有效。
平躺人體檢測具有很好的學術研究和應用價值。針對平躺人體錶觀姿態受透視變換的影響較大和樣本收集難這兩箇問題展開研究：首先建立併公開瞭廈門大學平躺人體檢測數據集，該數據集的人體包含各種姿態、視角、衣著顏色、複雜揹景等情況；其次提齣瞭一種基于透視變換的樣本擴充方法：根據透視變換的攝像機模型將3x3的透視變換矩陣進行分解，使得該矩陣最終隻與兩箇變化範圍確定的鏇轉角相關，方便瞭採樣算法的設計，有效地增加瞭訓練樣本的數目和樣本錶觀的多樣性，在廈門大學平躺人體數據集上的實驗結果錶明該方法提高瞭分類器的性能：噹FPPW=10e-4時，漏檢率降低瞭12.4%；噹FPPW=10e-5時，漏檢率降低瞭5.75%。可見本文提齣的樣本擴充方法簡單有效。
평당인체검측구유흔호적학술연구화응용개치。침대평당인체표관자태수투시변환적영향교대화양본수집난저량개문제전개연구：수선건립병공개료하문대학평당인체검측수거집，해수거집적인체포함각충자태、시각、의착안색、복잡배경등정황；기차제출료일충기우투시변환적양본확충방법：근거투시변환적섭상궤모형장3x3적투시변환구진진행분해，사득해구진최종지여량개변화범위학정적선전각상관，방편료채양산법적설계，유효지증가료훈련양본적수목화양본표관적다양성，재하문대학평당인체수거집상적실험결과표명해방법제고료분류기적성능：당FPPW=10e-4시，루검솔강저료12.4%；당FPPW=10e-5시，루검솔강저료5.75%。가견본문제출적양본확충방법간단유효。
Lying pose human detection has great research and application value. It is a difficult task since the appearance of lying pose human has many variations due to the perspective transformation of imaging, and the collection of samples is laborious. We focus on dealing with these two issues. Firstly, a lying pose human dataset is built and released openly, which includes many factors that affect the appearance of human, such as pose, viewpoint, clothes texture, and cluttered background. Second, an example expanding method based on perspective transformation is proposed to enrich the training dataset. It decomposes the three by three homograph matrix into a matrix ultimately determined by only two rang fixed rotation angles, facilitating the design of example sampling algorithm. Experimental results on our newly constructed dataset demonstrates that, when compared with classifier trained using samples without expanding, classifier trained by our proposed method decreases the miss rate with 12.4% when the false positive per window (FPPW) is 10e-4 and even decreases the miss rate with 5.75%when the FPPW is 10e-5, which verifies the effectiveness of our proposed method.
