JP4188068B2 - Observation target class identification device - Google Patents

Description

を満たす値として与えられる。ここで、位置正規化画素数Ｎｃは、以下の式（１）に基づいて算出される。
【００３７】
【数２】

【００３８】
ただし、式（１）において、Ｎは総画素数を表し、圧縮画像データＤ２の一部として与えられるものとする。
このように、この発明の実施の形態３によれば、個々の画像の画素数などの情報が得られない状態であっても、位置正規化の指標となるパラメータ（位置正規化率Ｒｃ）を事前に与えることができ、この指標に基づいて位置正規化の基準となる位置正規化画素順位Ｎｃを自動的に決定することができる。したがって、効率のよい位置正規化が可能となる。
【００３９】
実施の形態４．
なお、上記実施の形態３（図５）では、位置正規化画素順位決定器５６を設け、位置正規化率Ｒｃに基づいて位置正規化画素順位Ｎｃを自動的に決定したが、位置正規化画素順位決定器５６に代えて位置正規化画素順位変更器を設け、位置正規化率Ｒｃの変更割合を示すパラメータに基づいて位置正規化画素順位Ｎｃを自動的に変更してもよい。
以下、図６を参照しながら、位置正規化画素順位変更器５７を設けたこの発明の実施の形態４について説明する。
図６において、前述（図１、図４、図５参照）と同様のものについては、前述と同一符号を付して、または、符号の後に「Ｃ」を付して詳述を省略する。
【００４０】
この場合、類識別装置１２Ｃにおいて、画素指定データ圧縮器２と位置正規化器５Ｃとの間には、位置正規化画素順位変更器５７が挿入されている。
位置正規化画素順位変更器５７は、位置正規化パラメータＰ１８に含まれる位置正規化画素順位Ｎｃと圧縮画像データＤ２とを入力情報として、位置正規化変更率Ｕｃ（位置正規化率Ｒｃの変更割合を示すパラメータ）を決定するとともに、位置正規化変更率Ｕｃに基づいて位置正規化画素順位Ｎｃを自動的に変更して位置正規化画素順位変更値Ｎｃ’を生成する。
また、位置正規化器５Ｃは、圧縮画像データＤ２および位置正規化画素順位変更値Ｎｃ’を入力情報として、学習用位置正規化データＤ５ａおよび判別対象位置正規化データＤ５ｂを生成する。
【００４１】
次に、図６に示したこの発明の実施の形態４による具体的な処理動作について説明する。
この場合、類識別装置１２Ｃにおいて、位置正規化画素順位変更器５７は、圧縮画像データＤ２を第１の入力情報とし、位置正規化パラメータＰ１８に含まれる位置正規化画素順位Ｎｃを第２の入力情報として、位置正規化率Ｒｃを算出した後、位置正規化画素順位Ｎｃを変更するためのパラメータとして位置正規化変更率Ｕｃを決定し、位置正規化率Ｒｃおよび位置正規化変更率Ｕｃに基づいて位置正規化画素順位Ｎｃを変更し、変更結果を位置正規化画素順位変更値Ｎｃ’として出力する。
【００４２】
以下、位置正規化画素順位変更器５７による具体的な位置正規化画素順位Ｎｃの変更方法について説明する。
まず、圧縮画像データＤ２の一部として与えられた総画素数Ｎと、位置正規化画素順位Ｎｃとを用いて、以下の式（２）により位置正規化率Ｒｃを算出する。
【００４３】
【数３】

【００４４】
次に、以下の式（３）により、位置正規化画素順位変更値Ｎｃ’を算出する。
【００４５】
【数４】

【００４６】
ただし、式（３）において、位置正規化変更率Ｕｃは、
【数５】

を満たす値であり、乱数などを用いてランダムに決定される。
【００４７】
なお、ここでは、位置正規化率Ｒｃを算出した後に、位置正規化画素順位変更値Ｎｃ’を算出したが、前述の実施の形態３（図５）で示したように位置正規化率Ｒｃが事前に与えられている場合に適用することもできる。
この場合、位置正規化率Ｒｃの算出処理を省略して、同様に位置正規化画素順位変更値Ｎｃ’を算出すればよい。
【００４８】
このように、この発明の実施の形態４によれば、位置正規化の基準となる強度レベルを表す位置正規化画素順位Ｎｃを、事前に与えられた指標に基づいて自動的に変更することにより、観測条件などによって変化する画像の質に応じた強度レベルでの位置の正規化が可能となる。
【００４９】
実施の形態５．
なお、上記実施の形態３（図５）では、位置正規化画素順位決定器５６において、位置正規化率Ｒｃに基づいて位置正規化画素順位Ｎｃを決定したが、圧縮パラメータＰ１６に含まれる圧縮基準画素数と、位置正規化パラメータＰ１８に含まれる圧縮率依存位置正規化率とを用いて、圧縮後のデータ量に対する比率により位置正規化画素順位Ｎｃを決定してもよい。
以下、図７を参照しながら、圧縮後のデータ量に対する比率により位置正規化画素順位Ｎｃを決定したこの発明の実施の形態５について説明する。
図７において、前述（図１、図４、図５参照）と同様のものについては、前述と同一符号を付して、または、符号の後に「Ｄ」を付して詳述を省略する。
【００５０】
この場合、類識別装置１２Ｄにおいて、画素指定データ圧縮器２と位置正規化器５Ｄとの間には、位置正規化画素順位決定器５６Ｄが挿入されている。
位置正規化画素順位決定器５６Ｄは、圧縮画像データＤ２と、圧縮パラメータＰ１６に含まれるデータ圧縮後の画素数を表す圧縮基準画素数Ｎｓと、位置正規化パラメータＰ１８に含まれる圧縮率依存位置正規化率Ｒｓｃとを入力情報として、位置正規化画素順位Ｎｃを決定する。
また、位置正規化器５Ｄは、圧縮画像データＤ２および位置正規化画素順位Ｎｃを入力情報として、学習用位置正規化データＤ５ａおよび判別対象位置正規化データＤ５ｂを生成する
【００５１】
次に、図７に示したこの発明の実施の形態５による具体的な処理動作について説明する。
この場合、類識別装置１２Ｄにおいて、位置正規化画素順位決定器５６Ｄは、画素指定データ圧縮器２からの圧縮画像データＤ２を第１の入力情報とし、圧縮パラメータファイル１６からの圧縮基準画素数Ｎｓ（データ圧縮後の画素数を表すパラメータ）を第２の入力情報とし、位置正規化パラメータファイル１８からの圧縮率依存位置正規化率Ｒｓｃ（圧縮基準画素数Ｎｓから位置正規化画素順位Ｎｃを算出するためのパラメータ）を第３の入力情報として、圧縮基準画素数Ｎｓおよび圧縮率依存位置正規化率Ｒｓｃに基づいて位置正規化画素順位Ｎｃを算出する。
【００５２】
以下、位置正規化画素順位決定器５６Ｄによる位置正規化画素順位の決定方法について説明する。
位置正規化画素順位Ｎｃは、以下の式（４）に基づいて算出される。
【００５３】
【数６】

【００５４】
ただし、式（４）において、圧縮率依存位置正規化率Ｒｓｃは、
【数７】

を満たす値として与えられる。
【００５５】
このように、この発明の実施の形態５によれば、圧縮基準画素数Ｎｓとして与えられる圧縮後の画素数Ｎｓ個のうち、事前にパラメータとして与えられている割合（圧縮率依存位置正規化率Ｒｓｃ）の画素を位置正規化に適用することにより、圧縮状況に見合った強度レベルでの位置の正規化が可能となる。
【００５６】
実施の形態６．
なお、上記実施の形態２（図４）では、位置正規化器５Ａ内に順位指定位置平均算出器５１を設け、位置正規化パラメータＰ１８として指定された強度順位以上の全て画素を用いて位置平均Ｍｃを算出したが、位置正規化器内にさらに位置平均算出画素選択器を設け、位置正規化パラメータＰ１８として指定された強度順位以上の画素の中から、位置平均算出に適用する画素を選択して、位置平均Ｍｃを算出してもよい。
以下、図８を参照しながら、位置正規化器５Ｅ内に位置平均算出画素選択器５３を設けたこの発明の実施の形態６について説明する。
図８において、前述（図４参照）と同様のものについては、前述と同一符号を付して、または、符号の後に「Ｅ」を付して詳述を省略する。
【００５７】
この場合、類識別装置１２Ｅ内の位置正規化器５Ｅは、位置平均算出画素選択器５３と、順位指定位置平均算出器５１Ｅと、位置移動器５２とにより構成されており、位置平均算出画素選択器５３および順位指定位置平均算出器５１Ｅにおいて、位置正規化パラメータＰ１８として指定された強度順位以上の画素の中から、位置平均算出に適用する画素を選択して位置平均Ｍｃを算出するようになっている。
【００５８】
すなわち、位置平均算出画素選択器５３は、位置正規化パラメータＰ１８に含まれる位置正規化画素順位Ｎｃおよび位置平均算出画素数Ｎｒｃと圧縮画像データＤ２とを入力情報として、位置正規化画素順位Ｎｃで示される画素の中から位置平均算出画素数Ｎｒｃの画素を選択して位置平均算出画素Ｉｃとする。
また、順位指定位置平均算出器５１Ｅは、圧縮画像データＤ２および位置平均算出画素Ｉｃを入力情報として位置平均Ｍｃを算出し、位置移動器５２は、位置平均Ｍｃおよび圧縮画像データＤ２を入力情報として学習用位置正規化データＤ５ａおよび判別対象位置正規化データＤ５ｂを生成する。
【００５９】
次に、図８に示したこの発明の実施の形態６による具体的な処理動作について説明する。
類識別装置１２Ｅ内の位置正規化器５Ｅにおいて、まず、位置平均算出画素選択器５３は、画素指定データ圧縮器２からの圧縮画像データＤ２を第１の入力情報とし、位置正規化パラメータＰ１８に含まれる位置正規化画素順位Ｎｃを第２の入力情報とし、位置正規化パラメータＰ１８に含まれる位置平均算出画素数Ｎｒｃ（位置平均Ｍｃの算出に用いられる画素数）を第３の入力情報として、各圧縮画像データＤ２に対し、全画素を強度の降順にソートした後、位置正規化画素順位Ｎｃで与えられる強度上位Ｎｃ番目までの画素の中から、位置平均算出画素数Ｎｒｃ個の画素を選択して、位置平均Ｍｃの算出に用いられる画素を、位置平均算出画素Ｉｃとして出力する。
このとき、位置平均算出画素選択器５３による具体的な選択方法としては、たとえば、乱数などを発生させて、ランダムに選択する方法がある。
【００６０】
以下、順位指定位置平均算出器５１Ｅは、画素指定データ圧縮器２からの圧縮画像データＤ２と、位置平均算出画素選択器５３からの位置平均算出画素Ｉｃとを入力情報として、位置平均算出画素Ｉｃの位置平均Ｍｃを算出する。
また、位置移動器５２は、位置平均Ｍｃおよび圧縮画像データＤ２に基づいて学習用位置正規化データＤ５ａおよび判別対象位置正規化データＤ５ｂを生成する。
【００６１】
このように、この発明の実施の形態６によれば、強度上位となる画素のうち、ランダムに選択した画素を使って位置平均Ｍｃを求めることにより、画素数が比較的多い観測データＤ０において、画素数を絞り込むために、位置平均算出対象となる画素の強度を極端に高くすることなどによって、画素強度を偏らせることなく、位置平均算出処理の計算量を削減することができる。
【００６２】
実施の形態７．
なお、上記実施の形態６（図８）では、強度上位Ｎｃ番目までの画素の中からランダムに位置平均算出画素Ｉｃを選択したが、強度上位Ｎｃ番目までの画素が形成する図形の形状に基づいて、位置平均算出画素Ｉｃを選択してもよい。
以下、図８とともに、図９の説明図を参照しながら、強度上位Ｎｃ番目までの画素が形成する図形の形状に基づいて、位置平均算出画素Ｉｃを選択したこの発明の実施の形態７について説明する。
図９においては、位置平均算出画素Ｉｃの別の選択方法として、強度上位Ｎｃ番目までの画素が形成する図形の形状に基づいて、位置平均算出画素Ｉｃ（０）〜Ｉｃ（６）を選択する方法が示されている。
【００６３】
この場合、位置平均算出画素選択器５３は、画素指定データ圧縮器２からの圧縮画像データＤ２を第１の入力情報とし、位置正規化パラメータＰ１８に含まれる位置正規化画素順位Ｎｃを第２の入力情報として、各圧縮画像データＤ２に対し、全画素を強度の降順にソートした後、強度上位Ｎｃ番目までの画素から形成される凸包の頂点に位置する画素（Ｎｃｈ個）を、位置平均算出画素Ｉｃとして出力する。
【００６４】
図９は強度上位Ｎｃ番目までの画素から形成される凸包の例を示しており、図９において、各「黒丸マーク」は、強度上位Ｎｃ番目までの画素を表し、これらの画素を囲む多角形は、凸包を表している。
この場合、位置平均算出画素Ｉｃは、Ｉｃ（０）〜Ｉｃ（６）で表される。
【００６５】
このように、この発明の実施の形態７によれば、強度上位の画素で構成される凸包の各頂点に位置する画素Ｉｃ（０）〜Ｉｃ（６）を、位置平均算出対象画素Ｉｃとして選択することにより、画素数が比較的多いデータにおいて、画素数を絞り込むために、位置平均算出対象となる画素の強度を極端に高くすることなどによって、画素強度を偏らせることなく、且つ強度上位である画素の分布領域全体を考慮して、位置平均算出処理の計算量を削減することができる。
【００６６】
実施の形態８．
なお、上記実施の形態２（図４）では、位置正規化器５Ａ内に順位指定位置平均算出器５１を設け、全画素を強度の降順でソートし、位置正規化パラメータＰ１８として指定された順位以上の画素の位置平均Ｍｃを算出したが、順位指定位置平均算出器５１に代えて位置基準決定器を設け、指定された強度順位以上の画素の最大位置および最小位置に基づいて位置正規化処理を実行してもよい。
以下、図１０を参照しながら、位置正規化器５Ｆ内に位置基準決定器５４を設けたこの発明の実施の形態８について説明する。
図１０において、前述（図４参照）と同様のものについては、前述と同一符号を付して、または、符号の後に「Ｆ」を付して詳述を省略する。
【００６７】
この場合、類識別装置１２Ｆ内の位置正規化器５Ｆは、位置基準決定器５４および位置移動器５２Ｆにより構成されており、位置正規化画素順位Ｎｃにより指定された強度順位以上の画素の最大位置および最小位置に基づいて、位置正規化処理を実行するようになっている。
位置基準決定器５４は、位置正規化パラメータＰ１８に含まれる位置正規化画素順位Ｎｃと圧縮画像データＤ２とを入力情報として、位置正規化のための基準位置を表す位置正規化基準Ｌｃを出力する。
また、位置移動器５２Ｆは、圧縮画像データＤ２および位置正規化基準Ｌｃを入力情報として、位置正規化基準Ｌｃが圧縮画像データＤ２の各中心位置となるように、圧縮画像データＤ２を構成する全画素の位置を移動して、学習用位置正規化データＤ５ａおよび判別対象位置正規化データＤ５ｂを出力する。
【００６８】
次に、図１１の説明図を参照しながら、図１０に示したこの発明の実施の形態８による具体的な処理動作について説明する。
図１１は位置正規化基準決定器５４の処理動作を示す説明図であり、横軸はｘ座標、縦軸はｙ座標を示している。
図１１において、観測データＤ０の全画素は、二次元（ｘ，ｙ）平面上の点とみなされており、「黒丸マーク」は、強度上位Ｎｃ画素を表し、「白丸マーク」は、位置正規化基準Ｌｃおよび各線分を定義する４個の点ＣＬｃ（０）〜ＣＬｃ（３）を表している。
【００６９】
また、強度上位Ｎｃ画素のｘ座標およびｙ座標の最小値、最大値は、それぞれ、（ｘｍｉｎ，ｘｍａｘ）、（ｙｍｉｎ，ｙｍａｘ）で表され、それらの組み合わせに基づいて、各点ＣＬｃ（０）〜ＣＬｃ（３）が定義されている。
すなわち、点ＣＬｃ（０）およびＣＬｃ（２）を結ぶ線分と、点ＣＬｃ（１）およびＣＬｃ（３）を結ぶ線分との交点が位置正規化基準Ｌｃとなっている。
【００７０】
図１０において、位置正規化器５Ｆ内の位置基準決定器５４は、画素指定データ圧縮器２からの圧縮画像データＤ２を第１の入力情報とし、位置正規化パラメータＰ１８に含まれる位置正規化画素順位Ｎｃを第２の入力情報として、各圧縮画像データＤ２に対し、全画素を強度の降順にソートした後、強度上位Ｎｃ番目までの圧縮画像データＤ２の全画素を二次元平面上の点とみなした場合に、位置正規化画素順位Ｎｃで定義される画素の各座標の最小値および最大値の組み合わせから、位置正規化基準算出点ＣＬｃ（位置正規化の基準算出に用いられる点）を定義し、これらの位置正規化基準算出点ＣＬｃを結んだ線分の交点（Ｃｐ）を、位置正規化のための基準位置を表す位置正規化基準Ｌｃとして出力する。
【００７１】
図１１において、強度上位Ｎｃ画素の（ｘ，ｙ）座標の最小値、最大値（ｘｍｉｎ，ｘｍａｘ，ｙｍｉｎ，ｙｍａｘ）の組み合わせから定義される各点ＣＬｃ（０）〜ＣＬｃ（３）は、以下の式（５）のように表される。
【００７２】
【数８】

【００７３】
続いて、位置移動器５２Ｆは、画素指定データ圧縮器２からの圧縮画像データＤ２を第１の入力情報とし、位置基準決定器５４からの位置正規化基準Ｌｃを第２の入力情報として、位置正規化基準Ｌｃが各圧縮画像データＤ２の中心位置となるように、圧縮画像データＤ２を構成する全画素の位置を移動し、画素の移動結果を学習用位置正規化データＤ５ａおよび判別対象位置正規化データＤ５ｂとして出力する。
【００７４】
このように、この発明の実施の形態８によれば、強度上位の画素の２次元座標上の最大、最小値の組み合わせからなる４点を結んだ線分の交点を、位置正規化の基準位置（位置正規化基準Ｌｃ）とすることにより、強度上位である画素のうち、一部の画素が局所的に偏って（集中して）分布している場合であっても、その偏りの影響を受けることなく、強度上位である画素全体の分布状況に見合った位置の正規化を行うことが可能になる。
【００７５】
実施の形態９．
なお、上記実施の形態２（図４）では、位置正規化器５Ａ内に順位指定位置平均算出器５１を設け、全画素を強度の降順でソートして、指定された順位以上の画素の位置平均Ｍｃを算出したが、順位指定位置平均算出器５１に代えて画素指定位置平均算出器を設け、画素ＩＤにより指定された複数画素の位置平均を算出してもよい。
以下、図１２を参照しながら、位置正規化器５Ｇ内に画素指定位置平均算出器５５を設けたこの発明の実施の形態９について説明する。
図１２において、前述（図４参照）と同様のものについては、前述と同一符号を付して、または、符号の後に「Ｇ」を付して詳述を省略する。
【００７６】
この場合、類識別装置１２Ｇ内の位置正規化器５Ｇは、画素指定位置平均算出器５５および位置移動器５２Ｇにより構成されており、画素指定位置平均算出器５５は、画素ＩＤにより指定された複数画素の位置平均Ｍｃを算出する。
すなわち、画素指定位置平均算出器５５は、位置正規化パラメータＰ１８に含まれる複数の画素位置を表す位置正規化画素番号Ｍｃｎと、圧縮画像データＤ２とを入力情報として、圧縮画像データＤ２に対し、位置正規化画素番号Ｍｃｎから位置平均Ｍｃを算出する。
また、位置移動器５２Ｇは、位置平均Ｍｃおよび圧縮画像データＤ２を入力情報として、位置平均Ｍｃが圧縮画像データＤ２の各中心位置となるように、圧縮画像データＤ２を構成する全画素の位置を移動して、学習用位置正規化データＤ５ａおよび判別対象位置正規化データＤ５ｂを出力する。
【００７７】
次に、図１２に示したこの発明の実施の形態９による具体的な処理動作について説明する。
位置正規化器５Ｇにおいて、まず、画素指定位置平均算出器５５は、位置正規化パラメータＰ１８として指定される位置正規化画素番号Ｍｃｎ（複数の画素位置を示す）を第１の入力情報とし、画素指定データ圧縮器２からの圧縮画像データＤ２を第２の入力情報として、圧縮画像データＤ２に対し、位置正規化画素番号Ｍｃｎで指定された複数の画素に対して位置平均Ｍｃを算出する。
続いて、位置移動器５２Ｇは、位置平均Ｍｃおよび圧縮画像データＤ２を入力情報として、位置平均Ｍｃが圧縮画像データＤ２の各中心位置となるように、圧縮画像データＤ２を構成する全画素の位置を移動して、学習用位置正規化データＤ５ａおよび判別対象位置正規化データＤ５ｂを出力する。
【００７８】
このように、この発明の実施の形態９によれば、画像の画素数が少ない場合など、ユーザが画素の分布状況を比較的容易に把握可能な場合に、直接、位置平均Ｍｃの算出を行う画素を指定して位置平均Ｍｃの算出に用いることにより、全画像を対象としたソート処理が不要となり、ユーザの保有情報を有効に利用した効率のよい位置合わせが可能となる。
【００７９】
【発明の効果】
以上のように、この発明によれば、種類が不明の３次元の目標をレーダで観測した画像からなる観測データと、目標の種類の候補となる候補目標の学習用画像データを生成するための３次元の３Ｄ形状データと、観測データを学習するための学習パラメータと、観測データおよび学習用画像データを圧縮するためのデータ圧縮基準となる画素を示す圧縮パラメータと、観測データの位置を正規化するための画素の順位を表す位置正規化パラメータとを入力情報として、観測データの種類判別結果を出力する観測目標の類識別装置であって、観測データおよび３Ｄ形状データを入力情報として、観測データに基づいて目標の進行方向および回転に関する運動パラメータ値を決定するとともに、候補目標の３Ｄ形状データおよび運動パラメータ値に基づいて、目標の種類の学習に用いられる学習用画像データを生成する学習用画像生成器と、観測データ、圧縮パラメータおよび学習用画像データを入力情報として、圧縮パラメータに基づいて観測データおよび学習用画像データを圧縮して圧縮画像データを生成する画素指定データ圧縮器と、位置正規化パラメータおよび圧縮画像データを入力情報として、圧縮画像データに対し、全画素を強度の降順にソート後、位置正規化パラメータで指定された順位の画素を基準位置として、基準位置が画像の中心位置となるように全画素を移動させ、位置合わせを行うことにより、全画素の移動結果である学習用位置正規化データおよび判別対象位置正規化データを生成する位置正規化器と、学習パラメータおよび学習用位置正規化データを入力情報として、学習用位置正規化データに基づく種類学習を行い、種類学習結果を生成する種類学習器と、判別対象位置正規化データおよび種類学習結果を入力情報として、種類学習結果に基づいて判別対象位置正規化データの種類判別を行い、種類判別結果を出力する種類判別器とを備えたので、レーダ観測画像データを対象に、ユーザによる強度分布などのデータ解析の負荷を削減しつつ、レーダ観測画像データの性質に即した位置ずれに強い識別を実現した観測目標の類識別装置が得られる効果がある。
【図面の簡単な説明】
【図１】 この発明の実施の形態１を示すブロック構成図である。
【図２】 この発明の実施の形態１による位置正規化処理を示す説明図である。
【図３】 この発明の実施の形態１による学習処理および判別処理を示す説明図である。
【図４】 この発明の実施の形態２を示すブロック構成図である。
【図５】 この発明の実施の形態３を示すブロック構成図である。
【図６】 この発明の実施の形態４を示すブロック構成図である。
【図７】 この発明の実施の形態５を示すブロック構成図である。
【図８】 この発明の実施の形態６を示すブロック構成図である。
【図９】 この発明の実施の形態７による位置平均算出画素の選択処理動作を示す説明図である。
【図１０】 この発明の実施の形態８を示すブロック構成図である。
【図１１】 この発明の実施の形態８による位置正規化基準の決定処理動作を示す説明図である。
【図１２】 この発明の実施の形態９を示すブロック構成図である。
【符号の説明】
１ 学習用画像生成器、２ 画素指定データ圧縮器、３ 種類学習器、４ 種類判別器、５、５Ａ〜５Ｇ 位置正規化器、１２、１２Ａ〜１２Ｇ 類識別装置、１３ ３Ｄ形状データファイル、１４ 観測データファイル、１５ 種類判別結果ファイル、１６ 圧縮パラメータファイル、１７ 学習パラメータファイル、１８ 位置正規化パラメータファイル、５１ 順位指定位置平均算出器、５２、５２Ｆ、５２Ｇ 位置移動器、５３ 位置平均算出画素選択器、５４ 位置基準決定器、５５ 画素指定位置平均算出器、５６、５６Ｄ 位置正規化画素順位決定器、５７ 位置正規化画素順位変更器、Ｄ０ 観測データ、Ｄ１ 学習用画像データ、Ｄ２ 圧縮画像データ、Ｄ３ 種類学習結果、Ｄ４ 種類判別結果、Ｄ５ａ 学習用位置正規化データ、Ｄ５ｂ 判別対象位置正規化データ、Ｄ１３３Ｄ形状データ、Ｉｃ 位置平均算出画素、Ｌｃ 位置正規化基準、Ｍｃ 位置平均、Ｍｃｎ 位置正規化画素番号、Ｎｃ 位置正規化画素順位、Ｎｃ’ 位置正規化画素順位変更値、Ｎｒｃ 位置平均算出画素数、Ｎｓ 圧縮基準画素数、Ｐ１６ 圧縮パラメータ、Ｐ１７ 学習パラメータ、Ｐ１８ 位置正規化パラメータ、Ｒｃ 位置正規化率、Ｒｓｃ 圧縮率依存位置正規化率、Ｖｓ 強度基準値。[0001]
BACKGROUND OF THE INVENTION
The present invention relates to an apparatus that solves a class identification problem for determining a three-dimensional target type based on image data observed by a radar, and more specifically, to obtain a three-dimensional target type as a solution. The present invention relates to a class identification device for observation targets.
[0002]
[Prior art]
2. Description of the Related Art Conventionally, there is an apparatus for classifying an observation target such as a ship or an aircraft that performs class identification of a three-dimensional target from image data observed by a radar (for example, see Patent Document 1).
In the slave apparatus, first, posture / motion parameter values that are parameters such as the traveling direction / posture / rotational motion of each candidate target are estimated from observation data including tracking information, and the 3D shape data and posture / motion of the candidate target are estimated. Generate dictionary image data from parameter values, and then perform data compression on the observation data and dictionary image data using the intensity parameter, dictionary image data, and observation data as input information, and shift the target position between the compressed image data Is corrected and the alignment data is output, and finally, the correlation between the alignment data of the dictionary image data and the alignment data of the observation data is calculated, and an image having the maximum correlation is detected. .
[0003]
According to the above-described conventional apparatus, for each image data, the user designates one characteristic point as an alignment parameter, and these points are moved to coincide with the center position of each image. In this case, the user must specify the alignment parameter that is the center position based on the intensity distribution status of the individual image data. Have difficulty.
[0004]
[Patent Document 1]
Japanese Patent Laid-Open No. 6-174838 (FIG. 1)
[0005]
[Problems to be solved by the invention]
As described above, in the conventional observation target class identification device, the user must designate the alignment parameter to be the center position based on the intensity distribution status of each image data. However, when a large amount of data is targeted, it is difficult to specify manually.
[0006]
The present invention has been made to solve the above-described problems. For radar observation image data, while reducing the load of data analysis such as intensity distribution by the user, it is suitable for the characteristics of the radar observation image data. An object of the present invention is to obtain an observation target type classifying device that realizes strong discrimination against the misalignment.
[0007]
[Means for Solving the Problems]
The observation target class identification device according to the present invention generates observation data composed of images obtained by observing a three-dimensional target of unknown type with a radar, and learning image data for candidate targets that are candidates for the target type. 3D shape data representing a three-dimensional shape, learning parameters for learning observation data, compression parameters indicating pixels serving as a data compression reference for compressing observation data and learning image data, and observation data An observation target type identification device that outputs a type discrimination result of observation data by using, as input information, a position normalization parameter that represents a pixel order for normalizing the position, and inputs the observation data and 3D shape data as input information And determining motion parameter values related to the direction and rotation of the target based on the observation data, and the 3D shape data of the candidate target and A learning image generator for generating learning image data used for learning of the target type based on the dynamic parameter value, and observation based on the compression parameter using the observation data, the compression parameter, and the learning image data as input information A pixel designation data compressor that compresses data and learning image data to generate compressed image data, and a position normalization parameter and compressed image data as input information, After all pixels are sorted in descending order of intensity for compressed image data, the pixels in the order specified by the position normalization parameter are used as reference positions, and all pixels are moved so that the reference position becomes the center position of the image. By doing the combination The position normalizer for generating the learning position normalization data and the discrimination target position normalization data, which are the pixel movement results, and the learning parameter and the learning position normalization data as input information to the learning position normalization data. Based on the type learning result, the type of the classification target position normalized data is determined based on the type learning result. And a type discriminator for outputting the type discrimination result.
[0008]
DETAILED DESCRIPTION OF THE INVENTION
Embodiment 1 FIG.
Hereinafter, the first embodiment of the present invention will be described in detail with reference to the drawings.
FIG. 1 is a block diagram showing a first embodiment of the present invention.
In FIG. 1, the class identification device 12 includes a learning image generator 1 that generates learning image data D1 and pixel designation data that generates compressed image data D2 in order to identify the type of the observed three-dimensional target. The position inserted between the compressor 2, the type learner 3 that generates the type learning result D3, the type discriminator 4 that generates the type discrimination result D4, and the pixel designation data compressor 2 and the type learner 3. And a normalizer 5.
A 3D shape data file 13, an observation data file 14, a type determination result file 15, a compression parameter file 16, a learning parameter file 17, and a position normalization parameter file 18 are connected to the class identification device 12.
[0009]
The 3D shape data file 13 outputs 3D shape data D13 representing a three-dimensional shape. As described later, the 3D shape data D13 is used in the learning image generator 1 to generate candidate target learning image data D1 that is a candidate of the target type.
The observation data file 14 outputs observation data D0 including an image (including tracking information) obtained by observing a three-dimensional target of unknown type with a radar.
The type determination result file 15 stores the type determination result D4 finally output from the class identification device 12.
[0010]
The compression parameter file 16 outputs a compression parameter P16 indicating a pixel serving as a data compression reference for compressing the observation data D0 and the learning image data D1.
The learning parameter file 17 outputs a learning parameter P17 for learning the observation data D0.
The position normalization parameter file 18 outputs a position normalization parameter P18 representing the order of pixels for normalizing the position of the observation data D0.
[0011]
The class identification device 12 determines the type of the observation data D0 using the observation data D0, 3D shape data D13, compression parameter P16, learning parameter P17, and position normalization parameter P18 from the files 13, 14, 16-18 as input information. A result D4 is generated, and the type discrimination result D4 is input to the type discrimination result file 15.
In the class identification device 12, the learning image generator 1 uses the observation data D0 and the 3D shape data D13 as input information to generate learning image data D1 used for learning of the target type to be determined.
The pixel designation data compressor 2 uses the observation data D0, the compression parameter P16, and the learning image data D1 as input information, compresses the observation data D0 and the learning image data D1 based on the compression parameter P16, and compresses the compressed image data D2. Generate.
[0012]
The position normalizer 5 uses the position normalization parameter P18 and the compressed image data D2 as input information, moves the positions of all the pixels constituting the compressed image data D2 based on the position normalization parameter P18, and moves all the pixels. As a result, learning position normalized data D5a and discrimination target position normalized data D5b are generated.
The type learning device 3 performs type learning based on the learning position normalization data D5a using the learning parameter P17 and the learning position normalized data D5a as input information, and generates a type learning result D3.
The type discriminator 4 uses the discrimination target position normalized data D5b and the type learning result D3 as input information, performs type discrimination of the discrimination target position normalized data D5b based on the type learning result D3, and outputs a type discrimination result D4. .
[0013]
Specifically, the learning image generator 1 in the class identification device 12 determines the motion parameter values related to the target traveling direction and rotation based on the observation data D0, and the candidate target 3D shape data D13 and Learning image data D1 is generated based on the exercise parameter value.
[0014]
Next, the class identification operation according to the first embodiment of the present invention shown in FIG. 1 will be described with reference to FIGS.
FIG. 2 is an explanatory diagram showing position normalization processing by the position normalizer 5.
In FIG. 2, the left diagram shows the image state before normalization processing, and the middle diagram shows the image state (position normalization after normalization processing so that the reference position L becomes the center position of the image after calculating the reference position L. The right image shows another image of the same kind of target (position normalized image 2) that has been similarly normalized.
FIG. 3 is an explanatory diagram showing learning processing and discrimination processing by the type learning device 3 and the type discriminator 4, where the horizontal axis corresponds to attribute 1 and the vertical axis corresponds to attribute 2.
[0015]
First, in FIG. 1, the learning image generator 1 in the class identification device 12 uses the observation data D0 (including target tracking information) from the observation data file 14 as the first input information, and the 3D shape data file 13 The learning image data D1 is generated using the 3D shape data D13 of the candidate target from the second input information.
At this time, parameters necessary for generating the learning image data D1 include a motion parameter value that is a parameter such as a traveling direction / rotational motion of each candidate target, and a posture parameter value that is a parameter related to the posture of the candidate target. .
Among these, the motion parameter value is determined from the observation data D0 by estimation calculation or the like.
[0016]
In addition, the posture parameter values are generated as a plurality of parameter values based on the posture parameter generation reference given for generating the learning image data D1. As a specific example, in the case of the angle parameter θ [°], if the parameter interval is given as “30” as the posture parameter generation reference, the parameter value becomes an interval of 30 degrees, and θ = 0, 30, 60, 90, 120, 150, 180 are generated.
[0017]
Based on the motion parameter value and the posture parameter value determined in this way, the learning image generator 1 generates learning image data D1 from the 3D shape data D13 and the motion parameter value of each candidate target.
Of the parameters required for learning, the distinction between those that are alternatively estimated or determined and those that are generated as multiple values depends on the estimated accuracy, the nature of the candidate target, the observation conditions, etc. The user may decide accordingly.
[0018]
Subsequently, the pixel designation data compressor 2 uses the compression parameter P16 from the compression parameter file 16 as the first input information, the learning image data D1 from the learning image generator 1 as the second input information, and observes The compressed image data D2a and D2b are generated using the observation data D0 from the data file 14 as third input information.
Specifically, with respect to the observation data D0 and the learning image data D1, based on the pixel ID specified as the compression parameter P16, the intensity of a pixel equal to or lower than the pixel intensity indicated by the pixel ID (or less than the pixel intensity) is set to “ Data is compressed by converting it to “0”, and this data compression result is output as compressed image data D2.
[0019]
Next, the position normalizer 5 uses the position normalization parameter P18 from the position normalization parameter file 18 as first input information, and the compressed image data D2 from the pixel designation data compressor 2 as second input information. The position of all the pixels constituting the compressed image data D2 is moved based on the position normalization parameter P18 representing the order of the pixels for position normalization, and the deviation of the target position between the compressed image data D2 is corrected. Position normalization is performed to generate learning position normalization data D5a and discrimination target position normalization data D5b, which are the movement results (position normalization results) of all pixels.
[0020]
At this time, as a specific position normalization method, for each compressed image data D2, after sorting all the pixels in descending order of intensity, the pixel in the order specified by the position normalization parameter P18 is used as the reference position, and the reference position There is a method of performing alignment by moving all the pixels so that becomes the center position of the image.
In FIG. 2, a small square represents an individual pixel, a fill pattern of each pixel represents pixel intensity, and the denser the fill pattern, the stronger the pixel intensity.
In the left view (image before normalization), the central view (position normalized image 1), and the right view (position normalized image 2) in FIG. 2, the portions surrounded by thick frames are pixels of two images. This is where the intensities match.
By performing the position normalization process as shown in FIG. 2, many pixel intensities of the two images coincide with each other, so that the similarity becomes clear.
[0021]
According to the above-described position normalization method, the user can specify the reference pixel position without specifying a specific alignment point based on the intensity distribution of the individual image data of the radar observation image data (observation data D0). It is possible to align the observation data D0 only by specifying the rank.
That is, the user can realize identification that is resistant to misalignment based on information that is relatively easily available, such as the number of pixels.
[0022]
Hereinafter, the type learning device 3 determines the type learning result D3 based on the learning parameter P17 and the learning position normalization data D5a. Generate a Then, the type discriminator 4 generates the type discrimination result D4 based on the type learning result D3 and the discrimination target position normalized data D5b.
At this time, as a processing method by the type learning device 3, for example, a linear discriminating method for determining a boundary line is applied.
The linear discriminant method is a method for obtaining a straight line as a boundary based on the distance and direction from all the teacher data. The distance is maximized and the direction of the straight line as seen from the teacher data is the same for each type. Determine a straight line.
[0023]
In FIG. 3, the left figure shows the types and boundary lines learned by the linear discrimination method, and the right figure shows the types and boundary lines discriminated by the linear discrimination method.
In the left diagram of FIG. 3, the white triangle mark is the “type 1” teacher data, and the black square mark is the “type 2” teacher data, and the boundary line is determined as a straight line in the left diagram by type learning.
[0024]
Finally, the type discriminator 4 uses the type learning result D3 from the type learner 3 as the first input information, and the discrimination target position normalization data D5b (position discrimination target position normalization data from the position normalizer 5). ) As the second input information, the type of the determination target position normalized data D5b is determined based on the type learning result D3, and the type determination result D4 is output. This type discrimination result D4 becomes the output of the class discrimination device 12 and is input to the type discrimination result file 15.
[0025]
At this time, the type discriminator 4 performs type discrimination based on the boundary line obtained as the type learning result D3 as shown in the right diagram in FIG.
That is, in the right diagram of FIG. 3, the region above the boundary line to which many “white triangle marks” belong is defined as the boundary region of “Type 1”, and conversely, the region below the boundary line to which many “black square marks” belong. The region is defined as a “type 2” boundary region, and the type of the determination target observation data D5b is determined depending on which region it belongs to.
In the right diagram of FIG. 3, “black inverted triangle mark” and “white rhombus mark” are observation data whose types are undetermined, and are identified as “type 1” and “type 2” by the boundary lines, respectively. .
[0026]
Type learning and type discrimination by the above method is basically a method of determining a boundary line when there are two types, but when discrimination is made for a large number of types, for example, a method called a tournament method Is adopted.
In the tournament method, type learning is performed with all types of combinations of two types, and after determining each boundary line (discrimination plane), the type discrimination of which type the discrimination target data is close to is the winning type. Then, the combination of a plurality of types is performed, and the type that has finally won is determined as the target type.
[0027]
As described above, according to the first embodiment of the present invention, the user designates the rank of the pixel as the reference position without designating a specific alignment point based on the intensity distribution of the individual image data. As a result, it is possible to align the radar observation image data, and it is possible to realize identification that is resistant to misalignment from relatively easily available information such as the number of pixels.
[0028]
Embodiment 2. FIG.
In the first embodiment (FIG. 1), the position normalizer 5 executes the position normalization using the pixel of the intensity rank designated as the position normalization parameter P18 as the reference position. May be provided with a rank designation position average calculator and a position mover, and position normalization may be executed based on an average of the positions of pixels having an intensity rank or higher designated as the position normalization parameter P18.
Hereinafter, a second embodiment of the present invention in which a rank designated position average calculator 51 and a position mover 52 are provided in the position normalizer 5A will be described with reference to FIG.
In FIG. 4, the same components as those described above (see FIG. 1) are denoted by the same reference numerals as those described above, or “A” is appended to the reference numerals and detailed description thereof is omitted.
[0029]
In this case, the position normalizer 5A in the class identification device 12A is composed of the rank designation position average calculator 51 and the position mover 52, and the pixels of the intensity rank or higher designated as the position normalization parameter P18. Position normalization is performed based on the average position.
The rank designation position average calculator 51 in the position normalizer 5A uses the position normalized pixel rank Nc and the compressed image data D2 included in the position normalization parameter P18 as input information to perform compression based on the position normalized pixel rank Nc. Based on the average value of the vertical and horizontal positions of the image data D2, the position average Mc is calculated.
The position mover 52 uses the position average Mc and the compressed image data D2 as input information, and moves the positions of all the pixels constituting the compressed image data D2 so that the position average Mc becomes each center position of the compressed image data D2. Thus, the learning position normalization data D5a and the discrimination target position normalization data D5b are output.
[0030]
Next, a specific processing operation of the second embodiment of the present invention shown in FIG. 4 will be described.
In this case, in the position normalizer 5A in the class identification device 12A, first, the rank designation position average calculator 51 uses the position normalized pixel rank Nc included in the position normalization parameter as the first input information, and designates the pixel. The compressed image data D2 from the data compressor 2 is used as the second input information, and all the pixels are sorted in descending order of intensity for each compressed image data D2, and then the highest intensity Nc-th given by the position normalized pixel rank Nc. The average value of the vertical and horizontal positions of the pixels is determined as the position average Mc.
[0031]
Subsequently, the position mover 52 uses the position average Mc from the rank designation position average calculator 51 and the compressed image data D2 from the pixel designation data compressor 2 as input information, and the position average Mc is the value of each compressed image data. The positions of all the pixels constituting the compressed image data are moved so as to be the center position, and the movement results of the pixels are output as learning position normalized data D5a and discrimination target position normalized data D5b.
Here, the pixels up to the Nc-th position normalized pixel rank (highest intensity) are set as the position average calculation target. However, for example, when there are a plurality of pixels of the designated rank, The position average Mc may be calculated as a reference position.
[0032]
As described above, according to the second embodiment of the present invention, by determining the reference position for alignment based on the position average Mc of the plurality of pixels in the position normalized pixel rank (higher intensity) Nc, the higher intensity Nc It is possible to perform alignment corresponding to the distribution status of the pixels.
[0033]
Embodiment 3 FIG.
In the second embodiment (FIG. 4), the position normalization pixel rank Nc is given in advance as the position normalization parameter P18. However, the position normalization rate Rc is given as a parameter representing the position normalization ratio, The position normalized pixel rank Nc may be automatically determined by a position normalized pixel rank determiner in the class identification device.
Hereinafter, a second embodiment of the present invention in which a position normalized pixel rank determining unit is provided will be described with reference to FIG.
In FIG. 5, the same components as those described above (see FIGS. 1 and 4) are denoted by the same reference numerals as those described above, or “B” is appended after the reference numerals and detailed description thereof is omitted.
[0034]
In this case, in the class identification device 12B, a position normalized pixel rank determiner 56 is inserted between the pixel designation data compressor 2 and the position normalizer 5B.
The position normalized pixel rank determining unit 56 determines the position normalized pixel rank Nc using the position normalization rate Rc and the compressed image data 2D included in the position normalization parameter P18 as input information.
Further, the position normalizer 5B uses the position normalized pixel rank Nc determined by the position normalized pixel rank determiner 56 and the compressed image data D2 as input information, and the learning position normalized data D5a and the discrimination target position normalization. Data D5b is generated.
[0035]
Next, a specific processing operation according to the third embodiment of the present invention shown in FIG. 5 will be described.
In this case, in the class identification device 12B, the position normalized pixel rank determiner 56 uses the compressed image data D2 from the pixel designation data compressor 2 as the first input information, and the position normalization included in the position normalization parameter P18. The position normalized pixel rank Nc is determined based on the compressed image data D2 and the position normalization ratio Rc using the rate Rc (parameter for determining the position normalized pixel rank Nc) as the second input information.
[0036]
At this time, a specific method for determining the position normalized pixel rank Nc is, for example, as follows.
First, the position normalization rate Rc is
[Expression 1]

Is given as a value that satisfies Here, the position normalized pixel number Nc is calculated based on the following equation (1).
[0037]
[Expression 2]

[0038]
However, in Expression (1), N represents the total number of pixels and is given as a part of the compressed image data D2.
As described above, according to the third embodiment of the present invention, the parameter (position normalization rate Rc) serving as an index for position normalization is set even when information such as the number of pixels of each image cannot be obtained. The position normalization pixel rank Nc that is a reference for position normalization can be automatically determined based on this index. Therefore, efficient position normalization is possible.
[0039]
Embodiment 4 FIG.
In the third embodiment (FIG. 5), the position normalized pixel rank determining unit 56 is provided and the position normalized pixel rank Nc is automatically determined based on the position normalization rate Rc. A position normalized pixel rank changer may be provided instead of the rank determiner 56, and the position normalized pixel rank Nc may be automatically changed based on a parameter indicating the change rate of the position normalization rate Rc.
Hereinafter, a fourth embodiment of the present invention in which a position normalized pixel rank changer 57 is provided will be described with reference to FIG.
In FIG. 6, the same components as those described above (see FIGS. 1, 4 and 5) are denoted by the same reference numerals as those described above, or by adding “C” after the reference numerals and detailed description thereof will be omitted.
[0040]
In this case, in the class identification device 12C, a position normalized pixel rank changer 57 is inserted between the pixel designation data compressor 2 and the position normalizer 5C.
The position normalization pixel rank changing unit 57 receives the position normalization pixel rank Nc and the compressed image data D2 included in the position normalization parameter P18 as input information, and changes the position normalization change ratio Uc (change ratio of position normalization ratio Rc) And a position normalized pixel rank change value Nc ′ is generated by automatically changing the position normalized pixel rank Nc based on the position normalization change rate Uc.
Further, the position normalizer 5C generates the learning position normalized data D5a and the determination target position normalized data D5b using the compressed image data D2 and the position normalized pixel rank change value Nc ′ as input information.
[0041]
Next, a specific processing operation according to the fourth embodiment of the present invention shown in FIG. 6 will be described.
In this case, in the class identification device 12C, the position normalized pixel rank changing unit 57 uses the compressed image data D2 as the first input information and the position normalized pixel rank Nc included in the position normalization parameter P18 as the second input information. As information, after calculating the position normalization rate Rc, the position normalization change rate Uc is determined as a parameter for changing the position normalization pixel rank Nc, and based on the position normalization rate Rc and the position normalization change rate Uc The position normalized pixel order Nc is changed, and the change result is output as a position normalized pixel order change value Nc ′.
[0042]
Hereinafter, a specific method of changing the position normalized pixel order Nc by the position normalized pixel order changing unit 57 will be described.
First, the position normalization rate Rc is calculated by the following equation (2) using the total number of pixels N given as a part of the compressed image data D2 and the position normalized pixel rank Nc.
[0043]
[Equation 3]

[0044]
Next, the position normalized pixel rank change value Nc ′ is calculated by the following equation (3).
[0045]
[Expression 4]

[0046]
However, in the expression (3), the position normalization change rate Uc is
[Equation 5]

Which is determined at random using random numbers or the like.
[0047]
Here, after calculating the position normalization rate Rc, the position normalization pixel rank change value Nc ′ is calculated. However, as shown in the third embodiment (FIG. 5), the position normalization rate Rc is It can also be applied when given in advance.
In this case, the process of calculating the position normalization rate Rc may be omitted, and the position normalized pixel rank change value Nc ′ may be calculated in the same manner.
[0048]
As described above, according to the fourth embodiment of the present invention, by automatically changing the position normalization pixel rank Nc representing the intensity level as a reference for position normalization based on the index given in advance. In addition, it is possible to normalize the position at an intensity level corresponding to the quality of the image that changes depending on the observation conditions.
[0049]
Embodiment 5 FIG.
In the third embodiment (FIG. 5), the position normalized pixel rank determining unit 56 determines the position normalized pixel rank Nc based on the position normalization rate Rc, but the compression standard included in the compression parameter P16 is used. Using the number of pixels and the compression rate dependent position normalization rate included in the position normalization parameter P18, the position normalized pixel rank Nc may be determined based on the ratio to the data amount after compression.
Hereinafter, a fifth embodiment of the present invention in which the position normalized pixel rank Nc is determined based on the ratio to the data amount after compression will be described with reference to FIG.
In FIG. 7, the same components as those described above (see FIGS. 1, 4 and 5) are denoted by the same reference numerals as those described above, or by “D” after the reference numerals and detailed description thereof is omitted.
[0050]
In this case, in the class identification device 12D, a position normalized pixel rank determiner 56D is inserted between the pixel designation data compressor 2 and the position normalizer 5D.
The position normalization pixel rank determination unit 56D includes the compressed image data D2, the compression reference pixel number Ns representing the number of pixels after data compression included in the compression parameter P16, and the compression rate dependent position normalization included in the position normalization parameter P18. The position normalized pixel rank Nc is determined using the conversion rate Rsc as input information.
Further, the position normalizer 5D generates the learning position normalized data D5a and the discrimination target position normalized data D5b using the compressed image data D2 and the position normalized pixel rank Nc as input information.
[0051]
Next, a specific processing operation according to the fifth embodiment of the present invention shown in FIG. 7 will be described.
In this case, in the class identification device 12D, the position normalized pixel rank determination unit 56D uses the compressed image data D2 from the pixel designation data compressor 2 as the first input information, and the compression reference pixel number Ns from the compression parameter file 16 (Parameter representing the number of pixels after data compression) is used as the second input information, and the compression rate dependent position normalization rate Rsc (the position normalized pixel rank Nc is calculated from the compression reference pixel number Ns) from the position normalization parameter file 18 The position normalized pixel rank Nc is calculated based on the compression reference pixel number Ns and the compression rate dependent position normalization rate Rsc.
[0052]
Hereinafter, a method of determining the position normalized pixel order by the position normalized pixel order determiner 56D will be described.
The position normalized pixel rank Nc is calculated based on the following equation (4).
[0053]
[Formula 6]

[0054]
However, in equation (4), the compression rate dependent position normalization rate Rsc is
[Expression 7]

Is given as a value that satisfies
[0055]
Thus, according to the fifth embodiment of the present invention, the ratio (compression ratio-dependent position normalization ratio) given in advance as a parameter out of the number of compressed pixels Ns given as the compression reference pixel number Ns. By applying the pixel of Rsc) to the position normalization, it is possible to normalize the position at an intensity level suitable for the compression state.
[0056]
Embodiment 6 FIG.
In the second embodiment (FIG. 4), the rank specified position average calculator 51 is provided in the position normalizer 5A, and the position average is calculated using all the pixels having the intensity rank or higher specified as the position normalization parameter P18. Although Mc has been calculated, a position average calculation pixel selector is further provided in the position normalizer, and a pixel to be applied to the position average calculation is selected from among pixels having an intensity ranking or higher designated as the position normalization parameter P18. Then, the position average Mc may be calculated.
Hereinafter, a sixth embodiment of the present invention in which the position average calculation pixel selector 53 is provided in the position normalizer 5E will be described with reference to FIG.
In FIG. 8, the same components as those described above (see FIG. 4) are denoted by the same reference numerals as those described above, or “E” is appended to the reference numerals and detailed description thereof is omitted.
[0057]
In this case, the position normalizer 5E in the class identification device 12E includes a position average calculation pixel selector 53, a rank designation position average calculator 51E, and a position mover 52. In the unit 53 and the rank designation position average calculator 51E, the pixel to be applied to the position average calculation is selected from the pixels having the intensity rank or higher designated as the position normalization parameter P18, and the position average Mc is calculated. ing.
[0058]
That is, the position average calculation pixel selector 53 uses the position normalization pixel rank Nc and the position average calculation pixel number Nrc included in the position normalization parameter P18 and the compressed image data D2 as input information, and uses the position normalization pixel rank Nc. A pixel having the position average calculation pixel number Nrc is selected from the displayed pixels, and the position average calculation pixel Ic is selected. To .
The rank designation position average calculator 51E calculates the position average Mc using the compressed image data D2 and the position average calculation pixel Ic as input information, and the position mover 52 uses the position average Mc and the compressed image data D2 as input information. Learning position normalization data D5a and discrimination target position normalization data D5b are generated.
[0059]
Next, a specific processing operation according to the sixth embodiment of the present invention shown in FIG. 8 will be described.
In the position normalizer 5E in the class identification device 12E, first, the position average calculation pixel selector 53 uses the compressed image data D2 from the pixel designation data compressor 2 as the first input information, and sets the position normalization parameter P18. The position normalization pixel rank Nc included is used as second input information, and the position average calculation pixel number Nrc (number of pixels used to calculate the position average Mc) included in the position normalization parameter P18 is used as third input information. After all the pixels are sorted in descending order of intensity for each compressed image data D2, a pixel having the position average calculated pixel number Nrc is selected from the pixels up to the highest intensity Nc given by the position normalized pixel rank Nc. Then, the pixel used for calculating the position average Mc is output as the position average calculation pixel Ic.
At this time, as a specific selection method by the position average calculation pixel selector 53, for example, there is a method of generating random numbers and selecting them randomly.
[0060]
Hereinafter, the rank designation position average calculator 51E receives the compressed image data D2 from the pixel designation data compressor 2 and the position average calculation pixel Ic from the position average calculation pixel selector 53 as input information, and inputs the position average calculation pixel Ic. The position average Mc is calculated.
Further, the position mover 52 generates learning position normalization data D5a and discrimination target position normalization data D5b based on the position average Mc and the compressed image data D2.
[0061]
As described above, according to the sixth embodiment of the present invention, by obtaining the position average Mc using randomly selected pixels among the pixels with higher intensity, in the observation data D0 having a relatively large number of pixels, In order to narrow down the number of pixels, the amount of calculation of the position average calculation process can be reduced without biasing the pixel intensity by, for example, extremely increasing the intensity of the pixel that is the position average calculation target.
[0062]
Embodiment 7 FIG.
In the sixth embodiment (FIG. 8), the position average calculation pixel Ic is randomly selected from the pixels up to the highest intensity Nc. However, based on the shape of the figure formed by the pixels up to the highest intensity Nc. Thus, the position average calculation pixel Ic may be selected.
Hereinafter, referring to the explanatory diagram of FIG. 9 together with FIG. 8, the seventh embodiment of the present invention in which the position average calculation pixel Ic is selected based on the shape of the figure formed by the pixels up to the highest Nc-th intensity will be described. To do.
In FIG. 9, as another method for selecting the position average calculation pixel Ic, the position average calculation pixels Ic (0) to Ic (6) are selected based on the shape of the figure formed by the pixels up to the Nc-th highest intensity. The method is shown.
[0063]
In this case, the position average calculation pixel selector 53 uses the compressed image data D2 from the pixel designation data compressor 2 as the first input information, and sets the position normalized pixel rank Nc included in the position normalization parameter P18 to the second value. As input information, for each compressed image data D2, after sorting all the pixels in descending order of intensity, the pixels (Nch) located at the vertices of the convex hull formed from the pixels with the highest intensity Nc-th are obtained as the position average The calculated pixel Ic is output.
[0064]
FIG. 9 shows an example of a convex hull formed from pixels up to the highest intensity Nc. In FIG. 9, each “black circle mark” represents a pixel up to the highest intensity Nc, and many pixels surrounding these pixels. The square represents a convex hull.
In this case, the position average calculation pixel Ic is represented by Ic (0) to Ic (6).
[0065]
As described above, according to the seventh embodiment of the present invention, the pixels Ic (0) to Ic (6) located at the vertices of the convex hull composed of pixels with higher intensity are used as the position average calculation target pixel Ic. By selecting, in the data with a relatively large number of pixels, in order to narrow down the number of pixels, the intensity of the pixels that are the position average calculation target is made extremely high without increasing the intensity of the pixels. The calculation amount of the position average calculation process can be reduced in consideration of the entire pixel distribution area.
[0066]
Embodiment 8 FIG.
In the second embodiment (FIG. 4), the rank designation position average calculator 51 is provided in the position normalizer 5A, and all the pixels are sorted in descending order of intensity, and the rank designated as the position normalization parameter P18. Although the above-described pixel position average Mc is calculated, a position reference determining unit is provided in place of the rank designation position average calculator 51, and a position normalization process is performed based on the maximum position and the minimum position of the pixels having the specified intensity rank or higher. May be executed.
Hereinafter, an eighth embodiment of the present invention in which a position reference determiner 54 is provided in the position normalizer 5F will be described with reference to FIG.
In FIG. 10, the same parts as those described above (see FIG. 4) are denoted by the same reference numerals as those described above, or “F” is added after the reference numerals, and detailed description thereof is omitted.
[0067]
In this case, the position normalizer 5F in the class identification device 12F is configured by the position reference determiner 54 and the position mover 52F, and the maximum position of the pixels equal to or higher than the intensity rank specified by the position normalized pixel rank Nc. The position normalization process is executed based on the minimum position.
The position reference determiner 54 outputs a position normalization reference Lc representing a reference position for position normalization using the position normalization pixel rank Nc and the compressed image data D2 included in the position normalization parameter P18 as input information. .
Further, the position mover 52F receives all the compressed image data D2 and the position normalization reference Lc as input information, and sets all the compressed image data D2 so that the position normalization reference Lc is the center position of the compressed image data D2. The position of the pixel is moved, and the learning position normalized data D5a and the discrimination target position normalized data D5b are output.
[0068]
Next, a specific processing operation according to the eighth embodiment of the present invention shown in FIG. 10 will be described with reference to the explanatory diagram of FIG.
FIG. 11 is an explanatory diagram showing the processing operation of the position normalization criterion determiner 54, where the horizontal axis indicates the x coordinate and the vertical axis indicates the y coordinate.
In FIG. 11, all the pixels of the observation data D0 are regarded as points on the two-dimensional (x, y) plane, the “black circle mark” represents the higher intensity Nc pixel, and the “white circle mark” represents the position normality. The four points CLc (0) to CLc (3) that define the conversion standard Lc and each line segment are shown.
[0069]
In addition, the minimum value and the maximum value of the x coordinate and y coordinate of the higher intensity Nc pixel are represented by (xmin, xmax) and (ymin, ymax), respectively, and based on the combination thereof, each point CLc (0) ~ CLc (3) is defined.
That is, the intersection of the line segment connecting the points CLc (0) and CLc (2) and the line segment connecting the points CLc (1) and CLc (3) is the position normalization reference Lc.
[0070]
In FIG. 10, the position reference determiner 54 in the position normalizer 5F uses the compressed image data D2 from the pixel designation data compressor 2 as the first input information, and the position normalized pixels included in the position normalization parameter P18. Using the rank Nc as the second input information, after all the pixels are sorted in descending order of intensity for each compressed image data D2, all the pixels of the compressed image data D2 up to the Nc-th highest intensity are represented as points on the two-dimensional plane. When considered, a position normalization reference calculation point CLc (a point used for position normalization reference calculation) is defined from a combination of the minimum value and the maximum value of each coordinate of the pixel defined by the position normalization pixel rank Nc. Then, the intersection (Cp) of the line segment connecting these position normalization reference calculation points CLc is output as the position normalization reference Lc representing the reference position for position normalization.
[0071]
In FIG. 11, it is defined from the combination of the minimum value and the maximum value (xmin, xmax, ymin, ymax) of the (x, y) coordinates of the higher intensity Nc pixel. Each point CLc (0) to CLc (3) are expressed as the following formula (5).
[0072]
[Equation 8]

[0073]
Subsequently, the position mover 52F uses the compressed image data D2 from the pixel designation data compressor 2 as first input information and the position normalization reference Lc from the position reference determiner 54 as second input information. The positions of all the pixels constituting the compressed image data D2 are moved so that the normalization reference Lc becomes the center position of each compressed image data D2, and the movement results of the pixels are converted into the learning position normalized data D5a and the discrimination target position normalization. Is output as digitized data D5b.
[0074]
As described above, according to the eighth embodiment of the present invention, the intersection point of the line segment connecting the four points consisting of the combination of the maximum and minimum values on the two-dimensional coordinates of the pixel with higher intensity is used as the reference position for position normalization. By using (position normalization reference Lc), even if some pixels are locally biased (concentrated) among the pixels having higher intensity, the influence of the bias is reduced. Without receiving it, it becomes possible to normalize the position corresponding to the distribution status of the entire pixel having higher intensity.
[0075]
Embodiment 9 FIG.
In the second embodiment (FIG. 4), the rank designation position average calculator 51 is provided in the position normalizer 5A, and all the pixels are sorted in descending order of the intensity, and the positions of the pixels that are higher than the designated rank. Although the average Mc is calculated, a pixel specified position average calculator may be provided instead of the rank specified position average calculator 51 to calculate the position average of a plurality of pixels specified by the pixel ID.
Hereinafter, the ninth embodiment of the present invention in which the pixel designated position average calculator 55 is provided in the position normalizer 5G will be described with reference to FIG.
In FIG. 12, the same components as those described above (see FIG. 4) are denoted by the same reference numerals as those described above, or “G” is appended to the reference numerals and detailed description thereof is omitted.
[0076]
In this case, the position normalizer 5G in the class identification device 12G includes a pixel designated position average calculator 55 and a position mover 52G, and the pixel designated position average calculator 55 includes a plurality of pixels designated by the pixel ID. The pixel position average Mc is calculated.
That is, the pixel designation position average calculator 55 uses the position normalized pixel numbers Mcn representing the plurality of pixel positions included in the position normalization parameter P18 and the compressed image data D2 as input information, and inputs the compressed image data D2 to the compressed image data D2. The position average Mc is calculated from the position normalized pixel number Mcn.
Further, the position mover 52G takes the position average Mc and the compressed image data D2 as input information, and sets the positions of all the pixels constituting the compressed image data D2 so that the position average Mc becomes each center position of the compressed image data D2. It moves and outputs the learning position normalized data D5a and the discrimination target position normalized data D5b.
[0077]
Next, a specific processing operation according to the ninth embodiment of the present invention shown in FIG. 12 will be described.
In the position normalizer 5G, first, the pixel designated position average calculator 55 uses the position normalized pixel number Mcn (indicating a plurality of pixel positions) designated as the position normalization parameter P18 as first input information, Using the compressed image data D2 from the designated data compressor 2 as second input information, the position average Mc is calculated for a plurality of pixels designated by the position normalized pixel number Mcn for the compressed image data D2.
Subsequently, the position mover 52G uses the position average Mc and the compressed image data D2 as input information, and the positions of all the pixels constituting the compressed image data D2 so that the position average Mc becomes the center position of the compressed image data D2. And the learning position normalized data D5a and the discrimination target position normalized data D5b are output.
[0078]
As described above, according to the ninth embodiment of the present invention, the position average Mc is directly calculated when the user can grasp the pixel distribution state relatively easily, such as when the number of pixels of the image is small. By specifying a pixel and using it for calculation of the position average Mc, sorting processing for all images is not necessary, and efficient alignment using the user's possessed information effectively becomes possible.
[0079]
【The invention's effect】
As described above, according to the present invention, it is possible to generate observation data including images obtained by observing a three-dimensional target of unknown type with a radar, and learning image data for candidate targets that are candidates for the target type. Normalize the three-dimensional 3D shape data, a learning parameter for learning observation data, a compression parameter indicating a pixel serving as a data compression reference for compressing the observation data and learning image data, and the position of the observation data An observation target class identification device that outputs, as input information, a position normalization parameter that represents the order of pixels to be used, and outputs observation data type discrimination results, using observation data and 3D shape data as input information 3D shape data and motion parameter values of the candidate target are determined based on A learning image generator for generating learning image data to be used for learning of the target type based on the observation parameters, the compression parameters, and the learning image data as input information. A pixel designation data compressor that compresses image data to generate compressed image data, and a position normalization parameter and compressed image data as input information, After all pixels are sorted in descending order of intensity for compressed image data, the pixels in the order specified by the position normalization parameter are used as reference positions, and all pixels are moved so that the reference position becomes the center position of the image. By doing the combination , A position normalizer for generating learning position normalized data and discrimination target position normalized data, which are the movement results of all pixels, and learning position normalized data using learning parameters and learning position normalized data as input information Type learning based on the type learning result, type learning device that generates the type learning result, classification target position normalized data and type learning result as input information, type discrimination of the classification target position normalized data based on the type learning result Since it has a type discriminator that outputs the type discrimination result, it is possible to shift the position of the radar observation image data according to the characteristics of the radar observation image data while reducing the load of data analysis such as intensity distribution by the user. This makes it possible to obtain a class identification device for observation targets that realizes strong classification.
[Brief description of the drawings]
FIG. 1 is a block configuration diagram showing a first embodiment of the present invention.
FIG. 2 is an explanatory diagram showing position normalization processing according to the first embodiment of the present invention.
FIG. 3 is an explanatory diagram showing learning processing and discrimination processing according to Embodiment 1 of the present invention.
FIG. 4 is a block configuration diagram showing a second embodiment of the present invention.
FIG. 5 is a block diagram showing a third embodiment of the present invention.
FIG. 6 is a block diagram showing a fourth embodiment of the present invention.
FIG. 7 is a block diagram showing a fifth embodiment of the present invention.
FIG. 8 is a block diagram showing a sixth embodiment of the present invention.
FIG. 9 is an explanatory diagram showing a position average calculation pixel selection processing operation according to Embodiment 7 of the present invention;
FIG. 10 is a block diagram showing an eighth embodiment of the invention.
FIG. 11 is an explanatory diagram showing a position normalization reference determination processing operation according to Embodiment 8 of the present invention;
FIG. 12 is a block diagram showing a ninth embodiment of the invention.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 Learning image generator, 2 Pixel designation data compressor, 3 types learning device, 4 types discriminator, 5, 5A-5G position normalizer, 12, 12A-12G class identification device, 13 3D shape data file, 14 Observation data file, 15 type discrimination result file, 16 compression parameter file, 17 learning parameter file, 18 position normalization parameter file, 51 rank specified position average calculator, 52, 52F, 52G position mover, 53 position average calculation pixel selection , 54 position reference determiner, 55 pixel designated position average calculator, 56, 56D position normalized pixel rank determiner, 57 position normalized pixel rank changer, D0 observation data, D1 learning image data, D2 compressed image data , D3 type learning result, D4 type discrimination result, D5a learning position normalization data, D5b discrimination target position positive Data, D133D shape data, Ic position average calculation pixel, Lc position normalization reference, Mc position average, Mcn position normalization pixel number, Nc position normalization pixel rank, Nc ′ position normalization pixel rank change value, Nrc position average Calculated pixel count, Ns compression reference pixel count, P16 compression parameter, P17 learning parameter, P18 position normalization parameter, Rc position normalization ratio, Rsc compression ratio dependent position normalization ratio, Vs intensity reference value.

Claims (9)

Observation data composed of images obtained by observing a three-dimensional target of unknown type with a radar, 3D shape data representing a three-dimensional shape for generating candidate target learning image data that is a candidate of the target type, and , A learning parameter for learning the observation data, a compression parameter indicating a pixel serving as a data compression reference for compressing the observation data and the learning image data, and a position for normalizing the position of the observation data An observation target class identification device that outputs a type discrimination result of the observation data, using as input information a position normalization parameter that represents the order of pixels,
Using the observation data and the 3D shape data as input information, a motion parameter value related to the direction and rotation of the target is determined based on the observation data, and based on the 3D shape data and the motion parameter value of the candidate target A learning image generator for generating the learning image data used for learning the target type;
A pixel designation data compressor that generates compressed image data by compressing the observation data and the learning image data based on the compression parameter, using the observation data, the compression parameter, and the learning image data as input information;
Using the position normalization parameter and the compressed image data as input information, after sorting all pixels in descending order of intensity for the compressed image data, using the pixels in the order specified by the position normalization parameter as a reference position, Position normalization that generates learning position normalization data and discrimination target position normalization data, which are the movement results of all pixels, by moving all the pixels so that the reference position becomes the center position of the image and performing alignment And
A type learner that performs type learning based on the learning position normalized data and generates a type learning result using the learning parameter and the learning position normalized data as input information;
A type discriminator that performs type discrimination of the discrimination target location normalization data based on the type learning result using the discrimination target position normalized data and the type learning result as input information, and outputs the type discrimination result. An apparatus for classifying observation targets characterized by the fact that The position normalizer includes a rank designation position average calculator and a position mover,
The rank-designated position average calculator uses the position normalized pixel rank included in the position normalization parameter and the compressed image data as input information for each vertical and horizontal position of the compressed image data based on the position normalized pixel rank. Calculate the position average based on the average value,
The position mover uses the position average and the compressed image data as input information, and moves the positions of all the pixels constituting the compressed image data so that the position average becomes each central position of the compressed image data. 2. The observation target class identification device according to claim 1, wherein the learning position normalization data and the discrimination target position normalization data are output. A position normalized pixel rank determiner inserted between the pixel designation data compressor and the position normalizer;
The position normalization pixel rank determiner determines a position normalization pixel rank using the position normalization rate included in the position normalization parameter and the compressed image data as input information;
The position normalizer generates the learning position normalized data and the discrimination target position normalized data using the compressed image data and the position normalized pixel rank as input information. A class identification device for the described observation target. A position normalization pixel rank changer inserted between the pixel designation data compressor and the position normalizer;
The position normalization pixel rank changer is a position normalization that is a parameter for changing the position normalization pixel rank, using the position normalization pixel rank included in the position normalization parameter and the compressed image data as input information. And determining a normalized change rate, and changing the position normalized pixel rank based on the position normalized change rate to generate a position normalized pixel rank change value,
The position normalizer generates the learning position normalized data and the discrimination target position normalized data using the compressed image data and the position normalized pixel rank change value as input information. The observation target type identification device according to 1. A position normalized pixel rank determiner inserted between the pixel designation data compressor and the position normalizer;
The position normalized pixel rank determiner includes the compressed image data, a compression reference pixel number indicating the number of pixels after data compression included in the compression parameter, and a compression rate dependent position normalization included in the position normalization parameter. Using the rate as input information, the position normalized pixel ranking is determined,
The position normalizer generates the learning position normalized data and the discrimination target position normalized data using the compressed image data and the position normalized pixel rank as input information. A class identification device for the described observation target. The position normalizer includes a position average calculation pixel selector, a rank designation position average calculator and a position mover,
The position average calculation pixel selector uses the position normalization pixel rank, the position average calculation pixel number included in the position normalization parameter, and the compressed image data as input information, and outputs the pixels indicated by the position normalization pixel rank. Select the pixel of the position average calculation pixel number from among the position average calculation pixels used for calculation of the position average described later,
The rank designation position average calculator calculates the position average based on the average value of the vertical and horizontal positions of the compressed image data, using the compressed image data and the position average calculation pixel as input information,
The position mover uses the position average and the compressed image data as input information, and moves the positions of all the pixels constituting the compressed image data so that the position average becomes each central position of the compressed image data. 2. The observation target class identification device according to claim 1, wherein the learning position normalization data and the discrimination target position normalization data are output.   The position average calculation pixel selector uses, as input information, the compressed image data and the position normalized pixel rank, pixels located at the vertices of a convex hull formed by pixels indicated by the position normalized pixel rank, 7. The observation target class identification device according to claim 6, wherein the average calculation pixel is used. The position normalizer includes a position reference determiner and a position mover;
The position reference determiner, when the position normalized pixel order included in the position normalization parameter and the compressed image data are input information, and all the pixels of the compressed image data are regarded as points on a two-dimensional plane. A position normalization reference calculation point used for position normalization reference calculation is defined from a combination of minimum value / maximum value of each coordinate of the pixel defined by the position normalization pixel order, and the position normalization reference calculation is performed. The intersection of the line segments connecting the points is output as a position normalization reference representing the reference position for position normalization,
The position mover receives the compressed image data and the position normalization reference as input information, and sets all the pixels constituting the compressed image data so that the position normalization reference is a center position of the compressed image data. The observation target class identification device according to claim 1, wherein the learning position normalization data and the discrimination target position normalization data are output by moving a position. The position normalizer includes a pixel designated position average calculator and a position mover,
The pixel designated position average calculator uses the position normalized pixel numbers representing a plurality of pixel positions included in the position normalization parameter and the compressed image data as input information, and the position normalization is performed on the compressed image data. Calculate the position average from the pixel number,
The position mover uses the position average and the compressed image data as input information, and moves the positions of all the pixels constituting the compressed image data so that the position average becomes each central position of the compressed image data. 2. The observation target class identification device according to claim 1, wherein the learning position normalization data and the discrimination target position normalization data are output.

Images