JP3753014B2 - Traffic flow anomaly detection apparatus and method - Google Patents

Description

【００４８】
ここでσ²h,Aは上流の車両感知器で観測される車高の分散、σ²h,Bは上流の車両感知器で観測される車高の分散である。
同様に、上流の車両感知器で車長ｌAと検出された車長ｌの車両が、下流の車両感知器で車長ｌBと観測される確率は、下記(2)式で表される。
【００４９】
【数２】

【００５０】
ここでσ²l,Aは上流の車両感知器で観測される車長の分散、σ²l,Bは上流の車両感知器で観測される車長の分散である。
(a)コスト評価基準として車高又は車長の差を用いる場合
上流の車両感知器で車高ｈAと検出された車両が、下流の車両感知器で車高ｈBと検出される事後確率密度Ｐh(ｈa,ｈb)は、下記(3)式で表される。
【００５１】
【数３】

【００５２】
同様に、上流の車両感知器で車長ｌAと検出された車両が、下流の車両感知器で車長ｌBと検出される事後確率密度Ｐl(ｌa,ｌb)は、下記(4)式で表される。
【００５３】
【数４】

【００５４】
σ²h,A＋σ²h,B＝σ²hと表し、σ²l,A＋σ²l,B＝σ²lと書くと、前記(3)(4)式は、それぞれ、(5)(6)式のように書き換えられる。
【００５５】
【数５】

【００５６】
【数６】

【００５７】
(b) コスト評価基準として車高又は車長の比を用いる場合
前記(5)(6)式に代えて、次の(7)(8)式を用いる。
【００５８】
【数７】

【００５９】
【数８】

【００６０】
以上に掲げた確率密度の自然対数を「マッチコスト」という。車高のみを用いる場合の、上流の車両感知器で車高ｈAと検出された車両と、下流の車両感知器で車高ｈBと検出された車両とのマッチコストｄ(ｈa,ｈb)は、
ｄ(ｈa,ｈb)＝ln［Ｐh(ｈa,ｈb)］ (9)
となる。
車長のみを用いる場合は、上流の車両感知器で車長ｌAと検出された車両と、下流の車両感知器で車長ｌBと検出された車両とのマッチコストｄ(ｌa,ｌb)は、
ｄ(ｌa,ｌb)＝ln［Ｐl(ｌa,ｌb)］ (10)
となる。
【００６１】
車高と車長を併用する場合は、上流の車両感知器で車高ｈA、車長ｌAと検出された車両と、下流の車両感知器で車高ｈB、車長ｌBと検出された車両とのマッチコストｄ(ｈa,ｌa;ｈb,ｌb)は、
ｄ(ｈa,ｌa;ｈb,ｌb)＝ln［Ｐh(ｈa,ｈb)］＋ln［Ｐl(ｌa,ｌb)］ (11)
となる。
判定部Ｃは、前記(9)(10)(11)式のいずれかによって、上流の車両感知器で感知した車両、下流の車両感知器で感知した車両の全組み合わせについて、マッチコストを算出する(ステップＵ３)。
【００６２】
次に、判定部Ｃは、「ギャップコスト」を取得する。このギャップコストは、車両感知器の特性に応じた定数として記憶されているものである。マッチコストとギャップコストとを総称して「スコア」という。
ギャップコストには、インターナルギャップコスト（ＩＧＣ）とイクスターナルギャップコスト（ＥＸＧＣ）との２種類がある。
ＩＧＣは、両地点の間で追い越しが発生してその順序が入れ替わってしまい、対応が付かないような場合に設定される。これは、車高、車長のいずれかが３σ以上離れることはなく、３σ以上離れるとそれは別の車両とみなすとの仮定に基づく。
【００６３】
車高のみを用いる場合、ＩＧＣの具体的な値を示すと、(12)式のようになる。

式中１／２を用いているのは、ギャップコストを用いる場合は２本の枝を使うことになるので、枝１本分のコストにするためである（以下同じ）。
車長のみを用いる場合、ＩＧＣの具体的な値を示すと、(13)式のようになる。
【００６４】

車高と車長を併用する場合は、ＩＧＣの具体的な値を示すと、(14)式のようになる。

ＥＸＧＣは、上流地点を通過した車両がまだ下流地点を通過せず、車両の対応が付けられない場合に設定する。このＥＸＧＣの値は、実際に、本装置設置後の初期設定等の時点で、目視などで最適なマッチングが得られていると確認された場合に、アライメントの最大スコアを算出し、その最大スコアに基づいて決定されるものである。
【００６５】
さらに、現実には対応し得ないと思われる車両間の対応スコアは−∞にしておくことも考えられる（例えば旅行時間が負になる、経験上推定される旅行時間と比べると、あまりにも短いあるいは長い、など）。
以上の各コストが得られると、マッチコスト又はギャップコストの和が最大となるように、上流の車両感知器で感知したｎ台の車両と、下流の車両感知器で感知したｍ台の車両との対応付けを行う。このため、２次元の文字列アライメント問題として定式化する。
【００６６】
上流地点を通過した車両をａ1，ａ2，ａ3，‥‥，ａnで表し、下流地点を通過した車両をｂ1，ｂ2，ｂ3，‥‥，ｂnで表す。行列（ａiｂj）(1<i<n,1<j<m)をアライメントと呼ぶ。２次元の文字列アライメント問題は、２次元格子状有向グラフ上の最長路問題に帰着できる（一般的な呼び方は「最短路問題」であるが、ここではコストの和の最長のパスを求めているので、「最長路問題」という）。
図６は、２次元格子状有向グラフを描いた図である。図６では、車両ａiと車両ｂjとを対応させる斜めの枝が実線で表されている。この枝長は、前述したマッチコストｄ（(9)〜(11)式のいずれか）に相当する。枠の内側の縦横の破線枝は、車両ａiと車両ｂjとの対応が付かない場合を表し、その枝長は、前述したＩＧＣに相当する。枠の外側の一点鎖線の枝は、対応する車両がない場合を表し、その枝長は、前述したＥＸＧＣに相当する。
【００６７】
図６の左上の点から、右下の点に至るコストの和の最長のパスが、車両のもっともらしい対応付けを示す解となる。この最長路問題は、動的計画（ＤＰ;Dynamic Programming）法で解くことができる(ステップＵ４)（下記［１］［２］参照）。
［１］D. Gusfield. "Aigorithms on Strings, Trees, and Sequences." Cambridge University Prass, 1997．
［２］S.B.Needleman and C.D.Wunsch."A general method applicable to the search for similarities in the amino acid sequences of two proteins." Journal of Molecular Biology,48, pp.443-453, 1970．
例えば、ｎ＝３，ｍ＝４とし、最長路問題を解いた結果、図７に示すような経路（太い実線）が得られたとする。この図７から、車両ａ1はｂ1に対応し、車両ａ2と車両ｂ2は対応せず（インターナルギャップ）、車両ａ3はｂ3に対応し、車両ｂ4に対応するものがない（イクスターナルギャップ）、ことが分かる。この原因は、車両ａ2とａ3は上流地点を通過した後入れ替わった、と考えられ、車両ｂ4は途中の交差点から進入してきた若しくは車線間移動により入ってきた、と考えられる。
【００６８】
以上の解析結果を出力する(ステップＵ５)。そして、出力された結果に基づいて、一致する車両同士に着目して、旅行時間を推定することができる。
ここで、ＥＸＧＣの値を決定する方法を説明する。ＥＸＧＣの値をｇとおく。
上流でｎ台車両が観測され、下流でｍ台の車両が観測されたとする。ただしｎ＜ｍとする。目視などにより、上流地点で観測された車両ａ1〜ａnは、下流地点において観測された車両ｂ1〜ｂnに対応していることが分かっているものとする。しかし、上流地点で比較的遅い時刻に観測された車両ａn+1〜ａｍは、下流地点の観測時点において、まだ下流に到達しておらず、対応するものがない。
【００６９】
このときのアライメントの、マッチング部分のスコアをＳ′（Ａ）、非マッチング部分のスコアをｇ・ｘ（Ａ）と書く。Ｓ′（Ａ）は車両感知器の実測値から算出される値である。ｇは求めたいＥＸＧＣの値であり、ｘ（Ａ）は非マッチング台数を示し、ｘ（Ａ）＝ｎ−ｍである。
ＥＸＧＣの値ｇを「最適なマッチングが得られている場合の、マッチング部分のスコアＳ′（Ａ）を、ｎ＋ｍ−ｘ（Ａ）」で割ったもの、と定義する。
【００７０】
ｇ＝Ｓ′（Ａ）／［ｎ＋ｍ−ｘ（Ａ）］ (15)
たとえば、ｎ＝４，ｍ＝３とした場合の２次元格子状有向グラフを描くと図８のようになる。図８において、マッチング部分のスコアＳ′（Ａ）、ＥＸＧＣの値ｇ・ｘ（Ａ）を示している。ただし、ｘ（Ａ）＝ｎ−ｍ＝４−３＝１となるので、図８でｇ・ｘ（Ａ）と示したものは、ｇそのものを示している。
この図８の例では、ＥＸＧＣの値ｇは、前記(15)に基づき、
ｇ＝Ｓ′（Ａ）／［４＋３−１］＝Ｓ′（Ａ）／６ (16)
となる。
【００７１】
実際には、上下流でマッチングが確認されている車両群を観測してＥＸＧＣの値ｇを求める、という処理を複数回行い、求められた複数のｇの平均をとり、この平均値を最終的にＥＸＧＣの値ｇとして決定すればよい。
前記のＩＧＣとＥＸＧＣとを使った実施形態では、最新時点側（観測時刻の遅いほう）の車両対応付け部分（ｍの最も大きいところのデータ）は、ＥＸＧＣの値に大きく依存してしまう。ＥＸＧＣの値は、前述したように統計的に求められる値なので、変動する。
【００７２】
そこで、時系列最新時点側のＥＸＧＣを必要としないようにアルゴリズムを拡張する方法を説明する。
前述したとおり、ＥＸＧＣの値ｇは、
ｇ＝Ｓ′（Ａ）／［ｎ＋ｍ−ｘ（Ａ）］ (15)
で表される。一方、トータルスコアＳ（Ａ，ｇ）は、
Ｓ（Ａ，ｇ）＝Ｓ′（Ａ）＋ｇ・ｘ（Ａ） (17)
で表される。(15)式を(17)式に代入すると、
Ｓ（Ａ，ｇ）
＝Ｓ′（Ａ）＋Ｓ′（Ａ）ｘ（Ａ）／［ｎ＋ｍ−ｘ（Ａ）］
＝（ｎ＋ｍ）Ｓ′（Ａ）／［ｎ＋ｍ−ｘ（Ａ）］ (18)
となる。ｎ＋ｍは一定なので、Ｓ′（Ａ）／［ｎ＋ｍ−ｘ（Ａ）］を最大にするようなパスが最適な解となる。
【００７３】
上に述べたことを図解すると、次のようになる。
図９は、時系列最新時点側のイクスターナルギャップを除いた２次元格子状有向グラフを描いた図である。この図９のグラフにおいて、上流側通過車両ａ1，‥‥，ａnの添え字をｉ（１≦ｉ≦ｎ）とし、最新時点側の対応点をｖiとする。各点ｖiでのアライメントの最大スコアを求め、そのスコアを（ｉ＋ｍ）−ｘ（Ａ）で割った値が最大になる点を最新時点側の対応点ｖi₀とする。
【００７４】
以上で、本発明の実施の形態を説明したが、本発明の実施は、前記の形態に限定されるものではない。例えば、本発明において、車長や車高のデータ以外に、カメラで車両の画像データを取得してアライメントに利用することができる。この場合、画像間のスコアを定める必要があるが、画像検索などの分野で考案されている「画像間の距離」、具体的には、マッチディスタンス（下記［３］参照）、ＥＭＤ(Earth Mover's Distance)（下記［４］参照）を利用することができる。
【００７５】
［３］M.Werman, S.Peieg, and A.Rosenfeld."A distance metric for multi-dimensional histgrams." Computer, Vision, Graphics, and Image Processing, 32, pp.328-336, 1985.
［４］Y.Rubner, C Tomasi, and L.J.Guibas."The Earth Mover's Distance as a Metric for Image Retrieval. "Technical Report STAN-CS-TN-98-86, Department of Computer Science, Stanford University, September 1998.
マッチディスタンスは、画像内の画素の色情報を利用した距離で、２つの画像ヒストグラムの累積ヒストグラム間のＬ1距離として与えられる。具体的には、各車両の後方より撮影した画像から、輝度のヒストグラムと色相のヒストグラムを作る。輝度のヒストグラムは、すべての画素を輝度より３２レベルに分け、度数は各レベルの画素数とする。色相のヒストグラムは色相を３０段階に分け、度数は、各段階に含まれる画素の彩度の総和とする。画像の距離は、輝度と色相のヒストグラムによる距離の加重和とした。
【００７６】
ＥＭＤは、マッチディスタンスを一般化した距離で、画素の色情報の他に位置情報も用いる。各画像の画素を色と位置の近さによってクラスタリングし（下記［５］参照）、それらの間の最小費用流として与えられる。今回は、各画像を１６クラスタにクラスタリングした。
［５］M.Inaba, H.Imai, and N.Katoh."Application of weighted Voronoi diagrams and randomization to variance-based k-clastering." In Proceedings of the 10th ACM Symposium on Computational Geometry, pp.332-339, 1994
車両ａiと車両ｂjとの画像間の距離をＤijとしたとき、アライメントのスコアＤtotalとして、前記(9)(10)(11)式のいずれかと、画像間の距離Ｄijとの線形和を採用する。
【００７７】
Ｄtotal＝ｄ(ｈa,ｌa;ｈb,ｌb) ＋λＤij (18)
ここで定数λは重み付け係数である。また、車長・車高のギャップコストをＧとし、実対応車両間の距離の最大値をＭとしたとき、総ギャップコストＧtotalは、
Ｇtotal＝Ｇ−λＭ／２ (19)
として与えられる。
【００７８】
また、前記の実施の形態では、車両観測地点は、２地点としていたが、これを任意の複数地点に拡張することも可能である。車両観測地点がｐ（ｐは２以上の整数）地点あれば、車両の対応付けをｐ次元アライメント問題に帰着させることが可能である。
以上のようにして、上流地点を通過したｎ台の車両と、下流地点を通過したｍ台の車両との対応付けが行われた。
【００７９】
次に、判定部Ｃの行う交通流の異常検知処理をフローチャート（図１０）を用いて説明する。この異常検知処理は、上流地点を通過した車両と、下流地点を通過した車両との対応付けをするのに同期して、行う。
判定部Ｃは、上下流地点での対応付け結果を入力する(ステップＶ１)。
そして、下流地点を通過した車両のうち、何割が上流地点を通過した車両と対応付けされているのかを調べる。この割合をマッチング率という。例えば、下流地点を通過したｍ台の車両のうち、ｐ台が、上流地点を通過した車両と対応するとすれば、マッチング率は、ｐ／ｍとなる。
【００８０】
ここで、旅行時間を測定し、旅行時間が比較的短い場合、つまり自由流に近い場合、しきい値を比較的小さくし、旅行時間が比較的長い場合、つまり渋滞の場合、しきい値を比較的大きくする(ステップＶ２〜Ｖ４)。 判定部Ｃは、マッチング率としきい値とを比較し(ステップＶ２)、しきい値よりも低ければ、上下流地点の間で突発事象が発生したと判断する(ステップＶ４)。しきい値より高ければ、さらに、過去（例えば１回前）の上下流地点を通過した車両の対応付で得られたマッチング率がしきい値よりも高かったかどうかを調べ、高ければ異常なしと判断する。低ければ、突発事象が続いていると判断する。
【００８１】
このように、マッチング率が２回以上連続してしきい値よりも高い場合に、異常なしと判断するのは、突発事象が起こっても偶然マッチングした場合における誤った判断を避けるためである。
前記しきい値は、実際に突発事象が発生した場合のマッチング率を調べて、適切な値に決定すればよい。
以上のようにして、マッチング率及び突発事象発生の判定結果は、後述する総合判定に持ち込まれる。
【００８２】
なお、マッチング率を算出する方法として、他の公知の方法を採用てもよい。例えば、車両の画像を撮影し、画像処理を用いて車両の特徴、例えば色やプレートナンバーを認識する方法を採用してもよい。
３．４ 車両の平均速度に現れる疎密波（スペクトル変化）を利用した交通流の異常検知（方法Ｄ）
判定部Ｄは、一定（例えば１分）間隔の時刻ｔ1，ｔ2，ｔ3，‥‥，ｔn，‥‥（代表するときは添え字nを使う）ごとに交通量Ｑ(tn)、速度Ｖ(tn)、時間占有率Ｏ(tn)を記録し、それらの周波数スペクトルを算出する。この算出は、公知のように、交通量Ｑ(tn)、速度Ｖ(tn)又は時間占有率Ｏ(tn)の平均値を求め、時間ｋずらした標本自己共分散を求め、標本スペクトル分布ｐjを求めればできる。
【００８３】
具体的には、交通量Ｑ(tn)、速度Ｖ(tn)又は時間占有率Ｏ(tn)のいずれか１つを選び、それをｙnと表記する。ｙnの平均値Ｅ（ｙn）をμとおく。
μ＝Ｅ（ｙn）＝（１／Ｎ）Σｙn（総和はn＝１からＮまでとる。Ｎは標本数）
標本自己共分散関数Ｃkは、
Ｃk＝（１／Ｎ）Σ（ｙn−μ）（ｙn-k−μ）
となる。ただし総和Σは、n＝k+1からＮまでとる。標本スペクトルｐjは、
ｐj＝ΣＣk exp(-2πiｋfj)
となる。総和Σは、k＝-N+1からN-1までとる。余弦関数で表すと、
ｐj＝Ｃ0＋２ΣＣk cos(2πｋfj)
となる。総和Σは、k＝1からN-1までとる。ｆjは周波数であり、
ｆj＝ｊ／Ｎ
ｊ＝０，１，‥‥，［Ｎ／２］（［］はガウスの記号）
である。
【００８４】
判定部Ｄは、標本スペクトル分布に基づいて、突発事象発生を判定する。
図１１は、突発事象の発生した日に算出された交通量Ｑ(tn)の１分おきの変動を示すグラフである。突発事象発生地点の下流５００ｍの測定結果を示す。
突発事象の起こった時刻８時００分よりも前、Ｑ(tn)は５０台／分くらいで安定している。しかし、突発事象の起こった時刻８時００分の直後は、Ｑ(tn)は著しく低下する。交通障害物が除去された時刻８時２０分後、Ｑ(tn)は、６０台／分くらいに上がり、しばらくすると、もとの定常的な値５０台／分近辺に戻る。
【００８５】
図１２は、３０分間標本（Ｎ＝３０）での標本スペクトル（パワースペクトル）ｐjを求め、その分布を３次元的に図示したグラフである。横軸に周波数と時刻をとっている。縦軸はパワースペクトルである。
図１３は、標本スペクトルｐjのピーク値の軌跡を時間を追ってプロットしたグラフである。
事故の起こった時刻８時００分以前は、ピーク値は、ほとんど現れていない。８時００分以後、ピーク値は上昇し、交通障害物が除去された時刻８時２０分以後もその近辺の値を保つ。
【００８６】
図１４は、標本スペクトルのピーク周波数（単位Ｈｚ）をプロットしたグラフである。突発事象が起こった時刻８時００分以前は、ピーク周波数は比較的高めの値をとるが、突発事象が起こった時刻８時００分以後は、急激に下がり、０．０５Ｈｚ以下となる。交通障害物が除去された後も低い値を保つ。
図１５は、判定部Ｄが行う突発事象発生を判定する処理を説明するフローチャートである。図１５では、交通量Ｑ(tk)の標本スペクトル分布を求めることを想定して説明する。
【００８７】
判定部Ｄは、この処理を、時刻ｔnごとに繰り返し行う。図１５に沿って説明すると、判定部Ｄは、標本スペクトル分布ｐj（ｊは周波数相当）を算出する(ステップＷ０)。判定部Ｄは、判定部Ｄに付属するメモリに記憶される評価値（スコアという）を０に初期化し(ステップＷ１)、前回の突発事象発生判定フラグ又は交通状態注意判定フラグをチェックする(ステップＷ２)。
前回判定フラグオンであれば、前回のパワースペクトルの平均値を基準値として使用し(ステップＷ３)、一定値αをスコアに加算する(ステップＷ４)。
【００８８】
次に、スペクトルのピーク値（図１３参照）が過去３０分間のどのピーク値よりも大きいかどうか判定する(ステップＷ５)。大きければ、一定値βをスコアに加算する(ステップＷ６)。
次に、スペクトルのピーク値の変化量（例えば微分値）が過去３０分の平均値より大きいかどうか判定する(ステップＷ７)。大きければ、一定値γをスコアに加算する(ステップＷ８)。
【００８９】
次に、スペクトルのピーク周波数がしきい値（例えば０．１Ｈｚ）より小さいかどうか判定する(ステップＷ９)。小さければ一定値εをスコアに加算する(ステップＷ１０)。
前記スコアに加算するのは、瞬時的な誤検出を防止するためである。前記一定値α、β、γ、εをどのようにとるかは、実際に突発事象をどの程度正確に検知できるか、運用する高速道路を統計的に調べた上で、自動的に決定することができる。
【００９０】
そして、スコアに基づき、突発事象発生を判定し(ステップＷ１１)、判定結果を出力する(ステップＷ１２)。スコア及び判定結果は、総合判定に用いられる。
なお、前記の実施形態では、交通量Ｑ(tk)の標本スペクトル分布を求めていたが、速度Ｖ(tk)、時間占有率Ｏ(tk)の標本スペクトル分布を求めても、同様の処理により、突発事象発生判定が行える。また、これら以外にも、車頭時間間隔、道路区間の車両存在台数、車両の空間占有率、空間平均速度、車頭間隔距離の標本スペクトル分布を求めても、同様の処理により、突発事象発生判定が行える。
【００９１】
また、標本スペクトル分布ではパワースペクトルそのものの値やピーク周波数の値以外にも、スペクトル密度の値や位相特性を利用しても同様の効果が得られる。
３．５ 道路の交通量、車両の速度、画像データなどの測定値に基づく従来の交通流の異常検知（方法Ｅ）
判定部Ｅは、上流側の車両感知器３，車両感知器５での交通量又は車両の速度の測定値と、下流側での測定値との差を算出し、この差しきい値以上異なっているときに、当該区間での突発事象発生を判定し、その結果を総合判定に送る。
【００９２】
また判定部Ｅは、、道路に設置されたカメラによって道路の画像計測を行い、画像処理によって停止・低速車両の検出、車線変更車両の追跡等を行い、この結果に基づいて事故等の突発事象の発生を検知し、その結果を総合判定に送ってもよい（下記［６］参照）。
［６］山田、宮尾、酒井、西山、石下、根岸「トンネル内異常走行検知システムの開発」住友電気第１４５号、 pp.124 - 129, 1994年9月
また、前記以外に、ニューラルネットワークを利用した方法も採用可能である。
これらの公知の方法をまとめて「方法Ｅ」という。
【００９３】
４．総合判定
コンピュータ１１には、総合判定部２３が設けられている。総合判定部２３は、複数の判定アルゴリズムＡ〜Ｅの判定結果に基づいて、突発事象の発生尤度（確からしさ）を算出し、判定する。
４．１ 評価値の正規化
前記３．１〜３．５に述べた交通流の異常検知方法Ａ〜Ｅでは、それぞれ用語は違うが、評価値（図３）、相関（図４）、マッチング率（図１０）、スコア（図１５）、差といわれる値を算出し、それぞれ予めしきい値を設定しておき、これらの値がそれらのしきい値を超えたとき（相関の場合はしきい値より低いとき）に交通流の異常検知を行っていた。
【００９４】
「相関」の場合、他の方法と整合をとるためには、その逆数をとる必要がある。また、これらの値のとり得る範囲がばらばらであれば、総合判定ができないので、とり得る範囲を一定範囲に統一する必要がある。以後、０から１の範囲をとるようにこれらの値を正規化することとする。そして、正規化後の値を、「評価値」と統一的な名称で呼ぶことにする。
そして、方法Ａで算出された評価値をＰ_A，方法Ｂで算出された評価値をＰ_B，方法Ｃで算出された評価値をＰ_C，方法Ｄで算出された評価値をＰ_D，方法Ｅで算出された評価値をＰ_Eということにする。
【００９５】
これらの評価値Ｐ_A〜Ｐ_Eを用いた総合判定方法の例として、次の３つの判定方法を説明する。
以下では、公知の方法Ｅは説明の便宜上除外し、４つの方法Ａ〜方法Ｄを基本として説明するが、本発明の実施はこの形態に限られるものではなく、方法Ａ〜方法Ｄのいずれかを方法Ｅあるいは未知の方法で置き換えたり、方法Ａ〜方法Ｄに方法Ｅあるいは未知の方法を追加したりしても、本発明の実施は可能である。
【００９６】
４．２ 総合判定１（多数決方式）
この方法は、方法Ａ〜方法Ｄで異常検知した数に基づいて、総合判定を行う方法である。
図１６は、多数決方式を説明するためのフローチャートである。方法Ａ〜方法Ｄの処理がすべて終了していれば(ステップ(1))、３つ以上の方法で異常検知しているかどうか判定する(ステップ(2))。３つの以上の方法で異常検知していれば、突発事象の発生尤度が十分に高く「突発事象の発生」と判断する(ステップ(6))。
【００９７】
２つ以下の方法でしか異常検知していなければ、各方法で算出された評価値を加算する(ステップ(3))。この加算値Ｐ_A＋Ｐ_B＋Ｐ_C＋Ｐ_Dがしきい値を超えていれば、突発事象の発生尤度が中程度に高く、「突発事象の発生の可能性が高い注意状態」と判断する(ステップ(7))。なお、前記しきい値が高すぎると検知漏れが多くなり、しきい値が低すぎると誤検知が増えるので、しきい値が適切な値に設定されるようにしなければならない。このしきい値として、例えば、方法総数（いまの場合は４）の１／２倍という値を設定してもよい。
【００９８】
この加算値がしきい値を超えていなければ、突発事象の発生尤度が低く、「突発事象の発生なし」と判断する(ステップ(5))。
なお、以上の処理では、ステップ(2)において、４つある方法のうち３つ以上の方法で異常検知しているかどうか判定していたが、３という数に限られるものではなく、２又は４としてもよい。またステップ(4)において、しきい値を多数設け、突発事象の発生の尤度を段階的に複数個求めてもよい。
【００９９】
４．３ 総合判定２（重み付け方式）
方法Ａ〜方法Ｄで算出した評価値を、それぞれ重み付けして平均を求める方法である。それぞれ重み係数をα，β，γ，δとし、この重み付き平均値をＰ_totalと表示する。
Ｐ_total＝（αＰ_A＋βＰ_B＋γＰ_C＋δＰ_D）／（α＋β＋γ＋δ）
図１７は、重み付け方式を説明するためのフローチャートである。方法Ａ〜方法Ｄの処理がすべて終了していれば(ステップ(11))、上式に従ってＰ_totalを算出する。そしてこのＰ_totalを検知しきい値と比較し(ステップ(2))、検知しきい値を超えていれば突発事象の発生尤度が十分に高く「突発事象発生」と判断する(ステップ(15))。この検知しきい値が高すぎると検知漏れが多くなり、検知しきい値が低すぎると誤検知が増える。この検知しきい値は、後に図１８を用いて説明する検知漏れ率や誤検知率の実績に基づき、自動的に決定されるようにしてもよい。
【０１００】
検知しきい値を超えていなければ、このＰ_totalを注意しきい値と比較する(ステップ(14))。注意しきい値＜検知しきい値の関係がある。注意しきい値を超えていれば、突発事象の発生尤度が中程度に高く、「突発事象の発生の可能性が高い注意状態」と判断する(ステップ(16))。
注意しきい値を超えていなければ、突発事象の発生尤度が低く、「突発事象の発生なし」と判断する(ステップ(17))。
【０１０１】
ここで、前記重み係数α，β，γ，δの決定の仕方を説明する。この決定をする前提として、実際に交通計測データを集め、突発事象の発生時の交通計測データに基づいて、各方法Ａ〜方法Ｄで異常検知して、正しく検知したかどうかなどの実績を調べておく必要がある。
図１８は、この検知率等の記録方法を説明するためのフローチャートである。まず、交通計測データを常時集積する(ステップ(21))。実際に突発事象が発生したかことが分かると(ステップ(22)のYES)、発生時刻前後の交通計測データを参照し(ステップ(23))、方法Ａ〜方法Ｄの処理を行い（ステップ(24)）、それぞれの方法で評価値がしきい値を超えて交通流の異常検知を行っていたかどうか判断する。以上の処理を、突発事象が発生するたびに行う。
【０１０２】
この結果、突発事象全発生件数に対して正しく検知できた確率を「正検知率」、突発事象全発生件数に対して検知できなかった確率を「検知漏れ率」、総検知数に対して誤って検知した確率を「誤検知率」、突発事象が実際に発生してから検知するまでの時間を「検知遅れ時間」という(ステップ(25))。
総合判定部２３は、方法Ａ〜方法Ｄごとにこれらの値を、交通状態、曜日、季節、天候、時間帯別に分類し、記録している。次の表１は、記録内容を評価した一例である。
【０１０３】
【表１】

【０１０４】
総合判定部２３は、現在時刻の重み係数を決定する。
重み係数は、交通量Ｑ、速度Ｖ、占有率Ｏ、道路線形（カーブ、ジグザグ等）、曜日、時間帯、渋滞の程度、過去の検知実績（表１）などの関数とする。
図１９は、重み係数の決定処理を説明するためのフローチャートである。この処理は、リアルタイムで行う処理である。α〜δの初期値（例えば全部同一の値とする）に対して修正を施す。
【０１０５】
まず、検知の対象となる道路区間の道路線形による重みを加算する(ステップ(31))。例えば、ボトルネックとなりそうな道路線形であれば、方法Ａ（存在台数）の重み係数α、方法Ｂ（車線利用率）の重み係数βを上げる。
次に、曜日に基づいた重みを加算する(ステップ(32))。例えば現在が平日であれば、方法Ａ（存在台数）の重み係数α、方法Ｂ（車線利用率）の重み係数β、方法Ｃ（マッチング）の重み係数γを上げる。休日であれば、方法Ｄ（スペクトル）の重み係数δを上げる。
【０１０６】
次に、時間帯に基づいた重みを加算する(ステップ(33))。例えば朝であれば、方法Ａ（存在台数）の重み係数αを上げる。深夜であれば、方法Ｂ（車線利用率）の重み係数βを下げる。
次に、過去の実績に基づいた重みを加算する(ステップ(34))。例えば当該区間で検知率の高い方法の重み係数を上げる。
そして、今の交通状態（渋滞の程度）をチェックする(ステップ(35))。渋滞がなければ(ステップ(36)のNO)、方法Ｂ（車線利用率）の重み係数β、方法Ｄ（スペクトル）の重み係数δを上げる(ステップ(37))。
【０１０７】
渋滞があれば、渋滞初期かどうか判定し(ステップ(38))、渋滞初期であれば方法Ａ（存在台数）の重み係数αを上げる(ステップ(40))。渋滞中〜末期であれば方法Ｄ（スペクトル）の重み係数δを上げる(ステップ(39))。
以上のようにして、重み係数が自動的に決定される。
４．４ 総合判定３（検知順位方式）
この方式は、方法Ａ〜方法Ｄの検知遅れ時間に着目した総合判定方法である。
【０１０８】
簡単のため、２つの方法（方法Ａ〜方法Ｄのいずれでもよいが、ここでは方法Ａ、方法Ｄとする。）を例にとって説明する。
図２０は、過去の実績に基づいて、突発事象を方法Ａ、方法Ｄが同時に検知した割合、方法Ａが方法Ｄより早く検知した割合、方法Ａが方法Ｄより遅く検知した割合を示すグラフである。突発事象を方法Ａ、方法Ｄが同時に検知した割合４３％、方法Ａが方法Ｄより早く検知した割合３４％、方法Ａが方法Ｄより遅く検知した割合２３％となっている。この理由は、方法Ａが、突発事象発生後の渋滞初期に検知率が高い傾向があるのに対して、方法Ｄは突発事象発生後の渋滞中〜末期に検知率が高い傾向があることに基づくと考えられる。
【０１０９】
図２１(a)は、方法Ａ、方法Ｄが同時に検知した場合の、重みつけ平均による検知成功率、失敗率を示すグラフである。検知成功率、失敗率は例えば８９％，１１％となっている。図２１(b)は、方法Ａが方法Ｄより早く検知した場合の、重みつけ平均による検知成功率、失敗率を示すグラフである。検知成功率、失敗率は例えば１００％，０％となっている。図２１(c)は、方法Ｄが方法Ａより早く検知した場合の、重みつけ平均による検知成功率、失敗率を示すグラフである。検知成功率、失敗率は例えば８０％，２０％となっている。
【０１１０】
このように方法Ａが方法Ｄより早く検知した場合の検知成功率が１００％と高いので、本検知順位方式によれば、実際に方法Ａが方法Ｄより早く検知しているときは、無条件に突発事象発生と断定する。つまり異常発生の尤度を最大値に設定する。それ以外の場合に初めて、総合判定１又は総合判定２の方式で判定を行う。
図２２は、検知順位方式の判定方法を説明するためのフローチャートである。方法Ａ〜方法Ｄの処理がすべて終了し(ステップ(51))、少なくとも方法Ａ及び方法Ｄが突発事象発生を検知していれば(ステップ(52)のYES)、それぞれの方法Ａ，Ｄの検知時刻のデータを参照する(ステップ(53))。
【０１１１】
どの方法が早く検知したかを調べる(ステップ(54))。このとき、各方法の検知時刻が、同一の突発事象を検知しているとは考えられないほどかけ離れていれば、総合判定しても意味がない。従って、各方法の検知時間差が一定の時間あるいは一定の処理周期以内のみ、この総合判定３を行い、それ以外は、この判定をしないことが好ましい。
方法Ａが方法Ｂよりも早く検知していれば、尤度がもっとも高く、突発事象の発生を判定する(ステップ(55))。
【０１１２】
同時に検知したか、方法Ｂが方法Ａよりも早く検知しているときは、前述した総合判定1又は総合判定２の手法を用いて、突発事象の発生を判定する(ステップ(56))。突発事象の発生と判定されたときは(ステップ(57)のYES)、尤度が中程度であり、突発事象が発生した可能性があるとして、注意と判定する(ステップ(58))。突発事象の発生が判定されなかったときは、尤度は低く、平常状態と判定する(ステップ(59))。
【０１１３】
このように総合判定３は、方法Ａが方法Ｂよりも早く検知していれば、過去の傾向を重視して、前述した総合判定１又は総合判定２を行わずに、突発事象の発生を判定しているところに特徴がある。
４．５ 突発事象発生区間の特定
以上に説明した突発事象の発生が複数の区間で判定された場合、各区間における判定の尤度を比較して、もっとも尤度の高い区間を突発事象発生区間として特定することができる。
【０１１４】
以上のようにして突発事象の発生及びその発生区間が決定されると、交通管理センター１０は、可変表示板６，９に、「この先事故・止まれ」のような突発事象の発生を表示し、路側ビーコン７を通して車両に突発事象の発生を通知する。また、通信回線を通して関係機関１３や放送局１４に連絡する。また、交通状態注意の場合は、交通管理センター１０は、可変表示板６，９に「前方注意」のように運転者の注意を喚起するようなメッセージを表示し、路側ビーコン７を通して車両にも走行注意区間である旨を通知する。
【０１１５】
なお、すでに道路工事などが予定され、突発事象の発生が予想されている場合は、交通管理センター１０は、当該時刻に突発事象の発生を判定しても、この判定に基づいて可変表示板６，９に突発事象の発生を表示することはなく、関係機関１３や放送局１４に連絡することもない。
【０１１６】
【発明の効果】
以上のように本発明の交通流の異常検知装置又は方法によれば、各方式の検知結果を組み合わせることにより、各方式の欠点を補うことができるので、道路上の突発事象の発生をより精度よく検知することができる。
【図面の簡単な説明】
【図１】交通流の異常検知をするための交通流監視システムを示す概略図である。
【図２】交通管理センター１０内のコンピュータ１１の機能ブロック図である。
【図３】判定部Ａが行う突発事象発生を監視する処理を説明するためのフローチャートである。
【図４】判定部Ｂが行う突発事象発生を監視する処理を説明するためのフローチャートである。
【図５】マッチング判定部Ｃの行う車両群マッチング処理方法を説明するためのフローチャートである。
【図６】２次元格子状有向グラフを描いた図である。
【図７】ｎ＝３，ｍ＝４とし、最長路問題を解いた結果を示すための、２次元格子状有向グラフの図である。
【図８】ｎ＝４，ｍ＝３とし、車両ｂ4のみマッチングされていない場合の２次元格子状有向グラフである。
【図９】時系列最新時点側のイクスターナルギャップを必要としないアルゴリズムを説明するための、２次元格子状有向グラフを描いた図である。
【図１０】判定部Ｃが行う突発事象発生を監視する処理を説明するためのフローチャートである。
【図１１】実際に事故の発生した日に算出された交通量Ｑ(tn)の１分おきの変動を示すグラフである。
【図１２】標本スペクトル（パワースペクトル）ｐjを求め、その時間変化を図示したグラフである。
【図１３】標本スペクトルｐjを求め、そのスペクトルのピーク値をプロットしたグラフである。
【図１４】標本スペクトルのピーク周波数をプロットしたグラフである。
【図１５】判定部Ｄが行う突発事象発生を監視する処理を説明するためのフローチャートである。
【図１６】総合判定方法の１つである多数決方式を説明するためのフローチャートである。
【図１７】総合判定方法の１つである重み付け方式を説明するためのフローチャートである。
【図１８】過去の実績に基づく検知率等の記録方法を説明するためのフローチャートである。
【図１９】重み係数の決定処理を説明するためのフローチャートである。
【図２０】過去の実績に基づいて、突発事象を方法Ａ、方法Ｄが同時に検知した割合、方法Ａが方法Ｄより早く検知した割合、方法Ａが方法Ｄより遅く検知した割合を示すグラフである。
【図２１】 (a)は、方法Ａ、方法Ｄが同時に検知した場合の、重みつけ平均による検知成功率、失敗率を示すグラフである。 (b)は、方法Ｄが方法Ａより早く検知した場合の、重みつけ平均による検知成功率、失敗率を示すグラフである。(c)は、方法Ｄが方法Ａより早く検知した場合の、重みつけ平均による検知成功率、失敗率を示すグラフである。
【図２２】検知順位方式の判定方法を説明するためのフローチャートである。
【符号の説明】
１ 高速道路
２ 一般道路
３ 車両感知器
４ 一次処理装置
５ 車両感知器
６ 可変表示板
７ 路側ビーコン
９ 可変表示板
１０ 交通管理センター
１１ コンピュータ
１３ 関係機関
１４ 放送局
２１ 入力処理部
２２，Ａ〜Ｅ 判定部
２３ 総合判定部
２５ 出力処理部[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a traffic flow abnormality detection apparatus and method capable of collecting traffic measurement data by installing a vehicle detector or the like on a road, and detecting a traffic flow abnormality due to occurrence of a sudden event based on the traffic measurement data. Is.
[0002]
[Prior art]
When a sudden event such as a traffic accident or disaster occurs on the road, it is necessary to quickly detect an abnormality in the traffic flow based on the sudden event and notify the following vehicle or guide the following vehicle.
Conventionally, a camera is installed on a road and image processing is performed to detect an abnormality in a traffic flow (see JP-A-7-21488, JP-A-10-40490, etc.). It is expensive to install a camera over a wide area, and it is difficult to detect at night or in bad weather.
[0003]
In view of this, road traffic volume, vehicle speed, and the like are measured using vehicle detectors installed at various locations on the road, and traffic flow abnormalities are monitored based on these measured values.
According to this monitoring device, it is determined that an accident has occurred when the traffic has a small amount of traffic and the speed drops rapidly and the state continues for a certain period of time.
[0004]
[Problems to be solved by the invention]
However, in the monitoring device described above, since the determination is based on the traveling speed of the vehicle, there is a problem that it is difficult to distinguish when a sudden event occurs during a natural traffic jam, and the detection accuracy decreases.
Therefore, in order to solve the above-mentioned problem, the present inventor has detected a traffic flow abnormality detection algorithm (Japanese Patent Application No. 2000-199416) using time variation of the number of vehicles existing in a road section sandwiched between vehicle detectors. Traffic flow anomaly detection algorithm (Japanese Patent Application No. 2000-278352) focusing on uneven usage rate of lanes with roads, traffic by judging the degree of matching (matching) of the same vehicle group that passed two points above and below Flow anomaly detection algorithm (Japanese Patent Application No. 2000-289296), traffic flow anomaly detection algorithm (patent application 2000-) using the appearance of sparse waves (periodic disturbances, spectral changes) in the average traffic volume and vehicle speed No. 314139) and a patent application has been filed.
[0005]
If any one of the four traffic flow anomaly detection algorithms or the conventional traffic flow anomaly detection algorithm using changes in traffic volume or vehicle speed is used alone, an unexpected event on the road Can be detected with a certain degree of certainty.
However, none of these algorithms is universal, and depending on the conditions, detection accuracy may deteriorate. With only one algorithm, there is a high probability of false detection that a sudden event has not occurred, or there is a high probability of detection omission that it is determined that a sudden event has not occurred. . Some algorithms have a fast detection delay time and a slow algorithm.
[0006]
For example, in the traffic flow abnormality detection algorithm using the time variation of the number of vehicles existing in a road section, the detection accuracy is considered to decrease if the change in inflow and outflow of the vehicle decreases after a traffic jam occurs. In the traffic flow abnormality detection algorithm that pays attention to the bias in the utilization rate, if the sudden event occurrence point is too far from the vehicle detector, the detection accuracy is deteriorated. The traffic flow anomaly detection algorithm for judging matching of the same vehicle group and the traffic flow anomaly detection algorithm using spectrum change are effective in traffic jams, but when free running without traffic jams, it is more It is thought that the detection accuracy may deteriorate.
[0007]
Therefore, the inventor detects the occurrence of sudden events on the road more accurately by grasping changes in traffic conditions that appear when sudden events occur on the road from various aspects and comprehensively judging them. The present invention is proposed.
[0008]
[Means for Solving the Problems]
The traffic flow abnormality detection device according to the present invention combines an abnormality detection means for detecting an abnormality in a traffic flow by two or more different methods based on traffic measurement data, and a detection result of each method by the abnormality detection means. Comprehensive determination means for determining the occurrence of a flow abnormality, and the comprehensive determination means makes a comprehensive determination based on the fact that the detection time order detected by each method is in accordance with the order specific to the combination of the method ( Claim 1).
Depending on each method, there is a method that can be detected at an early stage and a method that is detected at a later stage. Therefore, if the certainty of detection is determined according to the time order detected by each method, the certainty of determination can be improved.
The two or more different methods include Two or more methods selected from (1) to (4) according to claim 2 based on traffic measurement data Can be raised (Claim 2).
[0014]
In addition, according to the present invention, it is preferable that the detection result in each method is obtained based on traffic measurement data before and after the occurrence of the sudden event, and accumulated as performance data (claim). 3 ). Thereby, evaluation based on the results of each method can be performed.
[0015]
The performance data may include one or more data among a positive detection rate, a detection omission rate, a false detection rate, and a detection delay time (claims). 4 ). These values are useful parameters for evaluating each method.
It is preferable that the traffic flow abnormality detection device of the present invention further includes information providing means for outputting the result of determination by the comprehensive determination means as abnormality information. 5 ). This is to prevent the accident from spreading by notifying the driver.
[0016]
in front The comprehensive determination means performs stepwise determination according to the magnitude (likelihood) of the value that is the basis of the determination, and the information providing means is abnormal depending on a result of the stepwise determination by the comprehensive determination means. Which changes the content of the information May be (Claims 6 ). Depending on the likelihood (certainty) of the occurrence of the sudden event, for example, by changing the contents of the information provision such as “future accident / stop”, “forward warning”, more appropriate information can be given to the driver.
[0017]
The comprehensive determination means may identify an abnormality occurrence section according to the magnitude of a value that is a basis of determination when an abnormality in traffic flow is detected in a plurality of road sections (claims). 7 ). By setting the section having the highest likelihood (certainty) of the occurrence of the sudden event as the abnormality occurrence section, it is possible to notify a subsequent driver or the like.
Further, the traffic flow abnormality detection method of the present invention (claims) 8,9 Are claims 1 and 2, respectively. 2 It is the method which concerns on the same invention as the abnormality detection apparatus of the following traffic flow.
[0018]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings, taking the traffic flow monitoring of an expressway as an example.
1. System configuration
FIG. 1 is a schematic diagram showing a traffic flow monitoring system for detecting an abnormality in traffic flow.
[0019]
On a highway 1 having two lanes, a two-loop embedded type vehicle detector 5 is installed for each lane at an interval. An ultrasonic vehicle detector 3 that measures the vehicle height from above the vehicle is also installed for each lane. The sections of the highway in which these vehicle detectors 3 and 5 are installed are displayed as sections 1, 2,..., I, (i is an integer of 2 or more). Let Li be the distance of each section. These vehicle detectors 3 and 5 and the camera are connected to the primary processing device 4, and the primary processing device 4 counts the number of passing vehicles, detects the vehicle speed, and the like.
[0020]
In addition, the highway 1 is provided with a variable display board 6 for notifying the vehicle of accident information, road surface information, and the like. A roadside beacon 7 that performs bidirectional communication with the vehicle is also provided.
Further, the general road 2 connected to the expressway 1 is provided with a variable display board 9 for notifying the vehicle that is about to enter the expressway 1 of accident information and road surface information of the expressway 1.
[0021]
A computer 11 inside the traffic management center 10 is connected to a primary processing device 4, a roadside beacon 7, a variable display board 6 and the like installed in each section through a wired communication network 12 (which may be a wireless communication network). Yes. In addition, the computer 11 is connected to related organizations 13 such as the Ministry of Construction, the National Police Agency and the Fire Department through a communication line, and is also connected to the broadcasting station 14 through the communication line.
In the above system example, a highway is assumed, but a general road may be used. Although a road with two lanes is assumed, the number of lanes is not limited to two, and may be one lane or three or more lanes.
[0022]
Further, instead of the plurality of embedded vehicle sensors 5, a Doppler vehicle sensor installed on the side of the road may be used. Alternatively, a television camera may be installed on the road to detect the number of passing vehicles, the vehicle height, the vehicle length, the passing speed, etc. by image processing.
All or some of the functions of the computer 11 described below are realized by the computer 11 executing a program recorded on a recording medium such as a program ROM.
[0023]
2. Traffic management center
FIG. 2 is a functional block diagram of the computer 11 inside the traffic management center 10.
A detection signal from the vehicle detector 5 is input to the input processing unit 21 of the computer 11 via the primary processing device 4. The input processing unit 21 determines the traffic volume (the number of passing vehicles per unit time) Q, the average speed V, the occupation rate O (a certain time T) based on the number of passing vehicles detected by the vehicle detector 5 and the detected amount of the vehicle speed. The sum Σtk of the time tk when the vehicle crosses the vehicle detector is divided by the time T: Σtk / T), the vehicle characteristic amount, and the like are calculated. The "average" speed is used to average the speed of each vehicle that has passed within a certain time. Hereinafter, the “average speed” is simply referred to as “speed”.
[0024]
Focusing on section i, the distance of section i is Li, the inflow traffic of the first lane to section i at time t is Q1, i (t), the inflow traffic of the second lane is Q2, i (t) Let Qi (t) be the inflow traffic volume into section i at time t, which is the sum of both the first lane and the second lane. Similarly, the first lane outflow traffic from section i is Q1, i + 1 (t), the second lane outflow traffic is Q2, i + 1 (t), and both the first lane and the second lane Qi + 1 (t) is the outflow traffic volume from section i at time t.
[0025]
Qi (t) = Q1, i (t) + Q2, i (t)
Qi + 1 (t) = Q1, i + 1 (t) + Q2, i + 1 (t)
Further, the inflow speed of the first lane to the section i at time t is V1, i (t), and the inflow speed of the second lane is V2, i (t). Similarly, the outflow speed of the first lane from section i is V1, i + 1 (t), and the outflow speed of the second lane is V2, i + 1 (t). The inflow speed into the section i at the time t obtained by weighted averaging of the first lane and the second lane with each traffic volume is Vi (t), and the outflow speed from the section i is Vi + 1 (t).
[0026]
Further, the occupation ratio of the vehicle in the first lane is O1 (t), and the occupation ratio of the vehicle in the second lane is O2 (t). Let O (t) be the occupation ratio of both lanes.
Further, the input processing unit 21 calculates a feature amount of the vehicle. That is, the maximum vehicle height of each vehicle is calculated based on the output of the vehicle detector 3, and the vehicle speed is measured based on the difference in output time between the two loops of the vehicle detector 5. The vehicle length is calculated based on Since the vehicle height and the vehicle length are calculated every time the vehicle passes, a data string of one or more vehicle heights and vehicle lengths is obtained for each lane.
[0027]
Although the traffic volume Q and the average speed V are calculated in the input processing unit 21 of the computer 11, these calculation processes may be performed in the primary processing device 4.
The detected data such as traffic volume, speed, occupation rate, vehicle height, vehicle length, etc. is referred to as “traffic measurement data”.
3. Individual judgment method
The computer 11 is provided with a determination unit 22, and the determination unit 22 includes determination units A to determine abnormalities in traffic flow using a plurality of determination algorithms (hereinafter referred to as method A to method E). A determination unit E is provided.
[0028]
3.1 Detection of traffic flow anomalies using time fluctuations in the number of vehicles in a road section (Method A)
As will be described below, the determination unit A calculates time fluctuations in the number of vehicles using different methods A1 and A2.
<Method A1>
The determination unit A calculates a difference ΔQi (t) between the inflow traffic volume Qi (t) and the outflow traffic volume Qi + 1 (t).
[0029]
ΔQi (t) = Qi (t) −Qi + 1 (t)
And the determination part A calculates | requires the amount of time fluctuations of the vehicle presence number Ei (t). ΔQi (t) itself represents the amount of time fluctuation of the vehicle existence number Ei (t).
The determination unit A records ΔQi (tk) for every time t1, t2, t3,..., Tk,. The variance over a period T (eg 10 minutes) is calculated. This variance is written as σ1 (tk).
[0030]
<Method A2>
The determination unit A divides the section i into the first half and the second half, and calculates the number of vehicles present in each. In the first half of the section, the number Ei1 (t) of the first half of the section is obtained using the inflow traffic volume Qi (t) and the inflow speed Vi (t).
Ei1 (t) = Qi (t) / Vi (t) · Li / 2
In the second half, the number Ei2 (t) of the second half of the section is obtained from the outflow traffic volume Qi + 1 (t) and the outflow speed Vi + 1 (t).
[0031]
Ei2 (t) = Qi + 1 (t) / Vi + 1 (t) · Li / 2
Then, the sum of both the existing number Ei1 (t) and Ei2 (t) is taken as the vehicle existing number Ei (t) in the section i.
Ei (t) = Ei1 (t) + Ei2 (t)
The determination unit A records Ei (tk) at each time tk and calculates the variance over the past period T in order to obtain the amount of time fluctuation of the number of vehicles existing Ei (t). This variance is written as σ2 (tk) (end of method A2).
[0032]
The determination unit A determines the occurrence of a sudden event based on the variances σ1 (tk) and σ2 (tk) acquired from the methods A1 and A2.
FIG. 3 is a flowchart for explaining the process of monitoring the occurrence of a sudden event performed by the determination unit A.
The determination unit A repeats this process every time tk. Explaining along FIG. 3, the determination unit A acquires the variances σ1 (tk) and σ2 (tk) (step S1), and whether σ1 (tk) and σ2 (tk) are each equal to or greater than the threshold value. Determination is made (steps S2, 4). If both are greater than or equal to the threshold value, a constant α is added to the evaluation value (step S3), and if only one is greater than or equal to the threshold value, a constant β (β <α) is added to the evaluation value (step S5). ). α and β are weighting coefficients of the methods A1 and A2. The reason for adding to the evaluation value is to prevent instantaneous erroneous detection.
[0033]
If σ1 (tk) and σ2 (tk) are both less than the threshold value, the evaluation value is reset to 0 (step S6).
It is determined whether the evaluation value is equal to or greater than a certain value (step S7). If the evaluation value is equal to or greater than the certain value, the determination unit A determines that a sudden event has occurred (step S8). If it is not greater than a certain value, the judgment is carried over at the next time tk + 1.
The calculated evaluation value and the determination result are sent to the comprehensive determination unit 23.
[0034]
3.2 Traffic flow anomaly detection focusing on lane usage rate bias (Method B)
The determination unit B determines the traffic volume Q1 (t) at regular time intervals (for example, 1 minute) every time t1, t2, t3,..., Tk,. , Q2 (t), speed V1 (t), V2 (t), occupancy O1 (t), O2 (t), and lane utilization rates r1 (t), r2 ( t) is calculated. Here, the lane utilization ratios r1 (t) and r2 (t) may be defined by the ratio of the traffic volume of the lane to the traffic volume of the entire lane, and the occupancy of the lane relative to the sum of the occupancy ratios of all lanes It may be defined by the ratio of rates. In the former case,
r1 (t) = Q1 (t) / Q (t),
r2 (t) = Q2 (t) / Q (t)
It is. In the latter case,
r1 (t) = O1 (t) / O (t),
r2 (t) = O2 (t) / O (t)
It is.
[0035]
The determination unit B calculates the average value of the lane utilization rate over the past period T (for example, 10 minutes). These average values are written as r1 and r2.
The determination unit B determines the occurrence of a sudden event according to the procedure shown in FIG.
FIG. 4 is a flowchart for explaining the process of monitoring the occurrence of a sudden event performed by the determination unit B. The following explanation is about the processing performed at a certain point where the vehicle detector 5 is installed. The same applies to each of the installation points of the vehicle detector 5 upstream and downstream. It is refused beforehand that processing is performed.
[0036]
The determination unit B repeats this process every time tk. If it demonstrates along FIG. 4, the determination part B will determine whether the traffic volume Q (t) is more than fixed (it may determine with occupation rate O (t)) (step T1). If the traffic volume Q (t) does not exceed a certain level, the determination is stopped. This is because if the traffic volume Q (t) is too low, the accuracy of the calculated lane utilization rate also decreases, so that a reliable determination of the occurrence of a sudden event cannot be made.
If the traffic volume Q (t) is above a certain level and the accuracy of the lane usage rate can be expected, the average lane usage rate r1 or r2 is calculated to determine whether the lane usage rate r1 or r2 is biased (step T2). ). The judgment criteria include whether r1 or r2 is close to 0.5 or whether the ratio of r1 to r2 is close to 1. For example,
r1 <0.4
Or
r1> 0.6
If so, it ’s determined that it ’s biased,
0.4 ≦ r1 ≦ 0.6
If so, it is determined that there is no bias. 0.6 and 0.4 are threshold values. If the threshold value is too close to 0.5, the chance that it is determined that the lane usage rate is biased increases, and the determination rate of occurrence of sudden events increases. As a result, it leads to excessive determination. If the threshold is too far from 0.5, the chance that the lane utilization rate is determined to be biased decreases, and the determination rate for occurrence of sudden events decreases. As a result, it is easy to overlook the occurrence of sudden events.
[0037]
Therefore, this threshold value records the lane utilization data before and after the actual occurrence of the sudden event on the road, and the occurrence of the sudden event on the road is most accurately detected by implementing the present invention. Choose a value that allows you to. In addition, data on lane usage under various conditions such as time of day, day of the week, day without special events, etc. is accumulated, and there is no optimum according to time of day, day of the week, season, weather, presence of events, etc. A threshold value may be set.
[0038]
When it is determined that the vehicle is biased, a correlation with the lane utilization rate detected by the vehicle detector 5 installed upstream is calculated (step T3). This correlation is given by, for example, the reciprocal of the difference between the lane usage rate r1 and the upstream lane usage rate r1 '(ABS represents an absolute value).
1 / ABS (r1-r1 ')
If the lane usage rate r1 and the upstream lane usage rate r1 'take a close value, the correlation increases. If the lane usage rate r1 and the upstream lane usage rate r1' take a distant value, the correlation Get smaller.
[0039]
Next, this correlation is compared with a certain reference value (step T4).
In this way, the correlation is compared with the reference value if the lane utilization rate r1 and the upstream lane utilization rate r1 ′ are separated (that is, the correlation is small) from the upstream point to the point. This is based on the fact that the utilization rate changes and it can be determined that a sudden event occurred during that time. As a “reference value”, a value that can detect the occurrence of a sudden event in a section with high accuracy is selected based on experience. Specifically, it is automatically selected from accumulated data.
[0040]
If the correlation with the upstream point is equal to or less than the reference value (NO in step T4), the correlation with the lane utilization rate detected by the vehicle detector 5 installed downstream is calculated (step T5).
If the correlation with the downstream point is also below the reference value (NO in step T6), it is determined that a sudden event has occurred near the point (step T7). This result is brought into the comprehensive determination described later.
If the correlation with the upstream point is equal to or higher than the reference value (when it is determined that there is a correlation) (YES in step T4), there is no change in the lane utilization rate from the upstream point to the point, and a sudden event occurs during that time It is difficult to think. Therefore, the determination process of the occurrence of the sudden event at the point is stopped (in this case, the occurrence of the sudden event may be detected in the sudden event occurrence monitoring process performed at the upstream point).
[0041]
If the correlation with the downstream point is equal to or greater than a certain reference value (when it is determined that there is a correlation) (YES in step T6), there is no change in the lane utilization rate from the point to the downstream point, and a sudden event in the meantime It is unlikely that this happened. Accordingly, also in this case, the determination process of the occurrence of the sudden event at the point is stopped (in this case, the occurrence of the sudden event may be detected in the sudden event occurrence monitoring process performed at the downstream point). .
[0042]
Note that the traffic flow abnormality detection method focusing on the lane utilization unevenness can be applied even on a three-lane road. In this case, the lane utilization rate r1 (t), r2 (t), r3 (t) of each lane may be defined by the ratio of the traffic volume of the lane to the traffic volume of the entire lane. You may define with the ratio of the occupation rate of the said lane with respect to. In the former case,
r1 (t) = Q1 (t) / Q (t),
r2 (t) = Q2 (t) / Q (t),
r3 (t) = Q3 (t) / Q (t),
Q (t) = Q1 (t) + Q2 (t) + Q3 (t)
It is. In the latter case,
r1 (t) = O1 (t) / O (t),
r2 (t) = O2 (t) / O (t),
r3 (t) = O3 (t) / O (t),
O (t) = O1 (t) + O2 (t) + O3 (t)
It is. Even on roads with 4 or more lanes, the lane utilization rate of each lane can be defined.
[0043]
3.3 Abnormality detection of traffic flow by judging matching of vehicle group (Method C)
A vehicle group matching processing method performed by the determination unit C will be described with reference to a flowchart (FIG. 5).
In addition to the method described below, various methods are known as the vehicle group matching processing method. In the following, as an example, a method in which vehicle matching is reduced to a p-dimensional alignment problem is performed. (Kobayashi et al. “Annual Traffic Flow Analysis Algorithm Based on Two-point Vehicle Observation Information” (published on March 21, 2000, 72nd Algorithm Study Group, Information Processing Society of Japan))
[0044]
In the following, attention is focused on only one lane. A vehicle that has entered the lane from another lane and a vehicle that has exited the lane from the lane are treated in the same manner as a vehicle that has entered the lane from an intersection or a vehicle that has exited the intersection from the lane.
Input the vehicle height and vehicle length data sequence detected by the vehicle sensors 3 and 5 upstream in the traveling direction (step U1), and input the vehicle height and vehicle length data sequence detected by the vehicle detectors 3 and 5 downstream in the traveling direction. (Step U2).
[0045]
The number of vehicles to be processed detected by the upstream vehicle sensor is n, the vehicle height data string is represented by hA1, hA2, hA3,..., HAn (represented by “hA” when representative), and the vehicle length data string. Is represented by lA1, lA2, lA3,..., LAn (represented by “lA” when representative). The number of vehicles to be processed detected by the downstream vehicle sensor is m, and the vehicle height data string is represented by hB1, hB2, hB3,..., HBm (represented by “hB” when represented), and the vehicle length data string. Is represented by lB1, lB2, lB3,..., LBm (represented by “lB” when representative).
[0046]
Assuming that the vehicle height h detected by the upstream vehicle detector is observed as the vehicle height hB at the downstream vehicle detector, the observation error of the vehicle detector follows a Gaussian distribution. Is expressed by the following equation (1).
[0047]
[Expression 1]

[0048]
Where σ ² h and A are the vehicle height dispersion observed by the upstream vehicle sensor, σ ² h and B are the vehicle height dispersion observed by the upstream vehicle sensor.
Similarly, the probability that a vehicle having a vehicle length l detected as a vehicle length lA by an upstream vehicle sensor is observed as a vehicle length lB by a downstream vehicle sensor is expressed by the following equation (2).
[0049]
[Expression 2]

[0050]
Where σ ² l, A is the dispersion of the vehicle length observed by the upstream vehicle sensor, σ ² l and B are vehicle length dispersion observed by the upstream vehicle detector.
(a) When the difference in vehicle height or vehicle length is used as a cost evaluation standard
The posterior probability density Ph (ha, hb) in which the vehicle detected as the vehicle height hA by the upstream vehicle detector and the vehicle height hB detected by the downstream vehicle detector is expressed by the following equation (3).
[0051]
[Equation 3]

[0052]
Similarly, the posterior probability density Pl (la, lb) in which the vehicle detected as the vehicle length lA by the upstream vehicle detector is detected as the vehicle length lB by the downstream vehicle detector is expressed by the following equation (4). Is done.
[0053]
[Expression 4]

[0054]
σ ² h, A + σ ² h, B = σ ² h and σ ² l, A + σ ² l, B = σ ² When written as l, the equations (3) and (4) are rewritten as equations (5) and (6), respectively.
[0055]
[Equation 5]

[0056]
[Formula 6]

[0057]
(b) When the vehicle height or vehicle length ratio is used as the cost evaluation standard
Instead of the above formulas (5) and (6), the following formulas (7) and (8) are used.
[0058]
[Expression 7]

[0059]
[Equation 8]

[0060]
The natural logarithm of the probability density listed above is called “match cost”. When only the vehicle height is used, the match cost d (ha, hb) between the vehicle detected as the vehicle height hA by the upstream vehicle detector and the vehicle detected as the vehicle height hB by the downstream vehicle detector is
d (ha, hb) = ln [Ph (ha, hb)] (9)
It becomes.
When only the vehicle length is used, the match cost d (la, lb) between the vehicle detected as the vehicle length lA by the upstream vehicle detector and the vehicle detected as the vehicle length 1B by the downstream vehicle detector is
d (la, lb) = ln [Pl (la, lb)] (10)
It becomes.
[0061]
When the vehicle height and the vehicle length are used in combination, the vehicle detected as the vehicle height hA and the vehicle length 1A by the upstream vehicle detector, and the vehicle detected as the vehicle height hB and the vehicle length 1B by the downstream vehicle detector Match cost d (ha, la; hb, lb) is
d (ha, la; hb, lb) = ln [Ph (ha, hb)] + ln [Pl (la, lb)] (11)
It becomes.
The determination unit C calculates the match cost for all combinations of the vehicle sensed by the upstream vehicle sensor and the vehicle sensed by the downstream vehicle sensor, according to any of the formulas (9), (10), and (11). (Step U3).
[0062]
Next, the determination unit C acquires “gap cost”. This gap cost is stored as a constant according to the characteristics of the vehicle detector. The match cost and the gap cost are collectively referred to as “score”.
There are two types of gap cost: internal gap cost (IGC) and external gap cost (EXGC).
The IGC is set when an overtaking occurs between the two points and the order is changed and no correspondence can be obtained. This is based on the assumption that either the vehicle height or the vehicle length is not separated by 3σ or more, and that it is regarded as another vehicle when separated by 3σ or more.
[0063]
When only the vehicle height is used, the specific value of IGC is expressed as shown in equation (12).

The reason why 1/2 is used in the equation is that when using the gap cost, two branches are used, so that the cost is equivalent to one branch (the same applies hereinafter).
When only the vehicle length is used, the specific value of IGC is shown as equation (13).
[0064]

In the case where the vehicle height and the vehicle length are used in combination, the specific value of IGC is shown in the equation (14).

EXGC is set when the vehicle that has passed through the upstream point has not yet passed through the downstream point and the vehicle cannot be dealt with. The EXGC value is calculated by calculating the maximum alignment score when it is confirmed that an optimum matching is obtained by visual inspection or the like at the time of initial setting after installation of the apparatus. It is determined based on
[0065]
Furthermore, the correspondence score between vehicles that cannot actually be handled may be set to −∞ (for example, the travel time becomes negative, which is too short compared with the travel time estimated from experience. Or long).
When each of the above costs is obtained, n vehicles sensed by the upstream vehicle sensor and m vehicles sensed by the downstream vehicle sensor so that the sum of the match cost or gap cost is maximized. Is associated. For this reason, it is formulated as a two-dimensional character string alignment problem.
[0066]
The vehicles that have passed the upstream point are represented by a1, a2, a3,..., An, and the vehicles that have passed the downstream point are represented by b1, b2, b3,. Matrix (aibj) (1 <i <n, 1 <j <m) is called alignment. The two-dimensional character string alignment problem can be reduced to the longest path problem on a two-dimensional grid directed graph (generally called the “shortest path problem”, but here the longest path of the sum of costs is obtained. (This is called the “longest road problem”).
FIG. 6 is a diagram depicting a two-dimensional grid-like directed graph. In FIG. 6, the diagonal branch which makes the vehicle ai correspond to the vehicle bj is represented by the solid line. This branch length corresponds to the above-mentioned match cost d (any one of the equations (9) to (11)). The vertical and horizontal broken lines inside the frame represent a case where the correspondence between the vehicle a i and the vehicle b j is not attached, and the length of the branch corresponds to the IGC described above. A one-dot chain line branch outside the frame represents a case where there is no corresponding vehicle, and the branch length corresponds to the above-described EXGC.
[0067]
The longest path of the sum of costs from the upper left point to the lower right point in FIG. 6 is a solution indicating a plausible association of vehicles. This longest path problem can be solved by a dynamic programming (DP) method (step U4) (see [1] and [2] below).
[1] D. Gusfield. “Aigorithms on Strings, Trees, and Sequences.” Cambridge University Prass, 1997.
[2] SBNeedleman and CDWunsch. “A general method applicable to the search for similarities in the amino acid sequences of two proteins.” Journal of Molecular Biology, 48, pp. 443-453, 1970.
For example, suppose that n = 3 and m = 4, and as a result of solving the longest path problem, a path (thick solid line) as shown in FIG. 7 is obtained. From FIG. 7, the vehicle a1 corresponds to b1, the vehicle a2 and the vehicle b2 do not correspond (internal gap), the vehicle a3 corresponds to b3, and there is nothing corresponding to the vehicle b4 (external gap). I understand. The cause is considered that the vehicles a2 and a3 have been switched after passing the upstream point, and the vehicle b4 has entered from an intermediate intersection or entered due to movement between lanes.
[0068]
The above analysis result is output (step U5). Based on the output result, the travel time can be estimated by paying attention to the matching vehicles.
Here, a method for determining the value of EXGC will be described. Let EXGC be g.
Assume that n vehicles are observed upstream and m vehicles are observed downstream. However, n <m. It is assumed that the vehicles a1 to an observed at the upstream point correspond to the vehicles b1 to bn observed at the downstream point by visual observation or the like. However, the vehicles an + 1 to am observed at a relatively late time at the upstream point have not yet reached the downstream at the time of observation at the downstream point, and there is no corresponding one.
[0069]
At this time, the score of the matching portion of the alignment is written as S ′ (A), and the score of the non-matching portion is written as g · x (A). S ′ (A) is a value calculated from the actual measurement value of the vehicle detector. g is the value of EXGC to be obtained, x (A) indicates the number of non-matching, and x (A) = n−m.
The value g of EXGC is defined as “the score S ′ (A) of the matching portion when optimum matching is obtained, divided by n + mx (A)”.
[0070]
g = S ′ (A) / [n + mx− (A)] (15)
For example, FIG. 8 shows a two-dimensional grid-like directed graph when n = 4 and m = 3. In FIG. 8, the score S ′ (A) of the matching portion and the EXGC value g · x (A) are shown. However, since x (A) = n−m = 4−3 = 1, what is indicated as g · x (A) in FIG. 8 indicates g itself.
In the example of FIG. 8, the value g of EXGC is based on (15) above.
g = S ′ (A) / [4 + 3-1] = S ′ (A) / 6 (16)
It becomes.
[0071]
Actually, the process of obtaining the EXGC value g by observing the vehicle group whose matching is confirmed in the upstream and downstream is performed a plurality of times, and the average of the obtained plurality of g is taken and this average value is finally obtained. The EXGC value g may be determined as follows.
In the embodiment using the IGC and EXGC described above, the vehicle association portion (data having the largest m) on the latest time point side (the later observation time) greatly depends on the value of EXGC. Since the value of EXGC is a value that is statistically obtained as described above, it fluctuates.
[0072]
Therefore, a method for extending the algorithm so as not to require EXGC on the time series latest time point side will be described.
As mentioned above, EXGC value g is
g = S ′ (A) / [n + mx− (A)] (15)
It is represented by On the other hand, the total score S (A, g) is
S (A, g) = S ′ (A) + g · x (A) (17)
It is represented by Substituting equation (15) into equation (17),
S (A, g)
= S '(A) + S' (A) x (A) / [n + mx (A)]
= (N + m) S '(A) / [n + mx (A)] (18)
It becomes. Since n + m is constant, a path that maximizes S ′ (A) / [n + mx− (A)] is the optimal solution.
[0073]
Illustrating what has been said above:
FIG. 9 is a diagram depicting a two-dimensional grid-like directed graph excluding the external gap on the time series latest time point side. In the graph of FIG. 9, the subscripts of the upstream passing vehicles a1,..., An are i (1 ≦ i ≦ n), and the corresponding point on the latest time point side is vi. The maximum score of the alignment at each point v i is obtained, and the point obtained by dividing the score by (i + m) −x (A) becomes the maximum corresponding point vi on the latest time point side. ₀ And
[0074]
Although the embodiments of the present invention have been described above, the embodiments of the present invention are not limited to the above-described embodiments. For example, in the present invention, in addition to vehicle length and vehicle height data, vehicle image data can be acquired by a camera and used for alignment. In this case, it is necessary to determine a score between images. “Distance between images” devised in the field of image search, specifically, match distance (see [3] below), EMD (Earth Mover's Distance) (see [4] below) can be used.
[0075]
[3] M. Werman, S. Peieg, and A. Rosenfeld. "A distance metric for multi-dimensional histgrams." Computer, Vision, Graphics, and Image Processing, 32, pp.328-336, 1985.
[4] Y.Rubner, C Tomasi, and LJGuibas. "The Earth Mover's Distance as a Metric for Image Retrieval." Technical Report STAN-CS-TN-98-86, Department of Computer Science, Stanford University, September 1998.
The match distance is a distance using the color information of the pixels in the image and is given as an L1 distance between the cumulative histograms of the two image histograms. Specifically, a luminance histogram and a hue histogram are created from images taken from the rear of each vehicle. In the luminance histogram, all pixels are divided into 32 levels from the luminance, and the frequency is the number of pixels at each level. The hue histogram divides the hue into 30 stages, and the frequency is the sum of the saturations of the pixels included in each stage. The image distance is the weighted sum of the distances based on the histogram of luminance and hue.
[0076]
EMD is a distance that generalizes match distance, and uses position information in addition to pixel color information. The pixels of each image are clustered by color and position proximity (see [5] below), given as the minimum cost flow between them. This time, each image was clustered into 16 clusters.
[5] M. Inaba, H. Imai, and N. Katoh. "Application of weighted Voronoi diagrams and randomization to variance-based k-clastering." In Proceedings of the 10th ACM Symposium on Computational Geometry, pp.332-339, 1994
When the distance between the images of the vehicle ai and the vehicle bj is Dij, a linear sum of any of the expressions (9), (10), and (11) and the distance Dij between the images is adopted as the alignment score Dtotal. .
[0077]
Dtotal = d (ha, la; hb, lb) + λDij (18)
Here, the constant λ is a weighting coefficient. In addition, when the gap cost of the vehicle length / height is G and the maximum distance between actual vehicles is M, the total gap cost Gtotal is
Gtotal = G-λM / 2 (19)
As given.
[0078]
In the above embodiment, the number of vehicle observation points is two. However, it is possible to extend this to any number of points. If the vehicle observation point is a p point (p is an integer of 2 or more), it is possible to reduce the vehicle association to the p-dimensional alignment problem.
As described above, the n vehicles that have passed the upstream point and the m vehicles that have passed the downstream point are associated with each other.
[0079]
Next, the traffic flow abnormality detection process performed by the determination unit C will be described with reference to a flowchart (FIG. 10). This abnormality detection process is performed in synchronization with the association between the vehicle that has passed the upstream point and the vehicle that has passed the downstream point.
The determination unit C inputs the association result at the upstream and downstream points (step V1).
Then, it is examined how many of the vehicles that have passed the downstream point are associated with the vehicles that have passed the upstream point. This ratio is called the matching rate. For example, out of m vehicles that have passed the downstream point, if the p vehicles correspond to the vehicles that have passed the upstream point, the matching rate will be p / m.
[0080]
Here, the travel time is measured, and if the travel time is relatively short, i.e., close to free flow, the threshold value is relatively small, and if the travel time is relatively long, i.e., in a traffic jam, Make it relatively large (steps V2 to V4). The determination unit C compares the matching rate with a threshold value (step V2), and if lower than the threshold value, determines that a sudden event has occurred between the upstream and downstream points (step V4). If it is higher than the threshold, it is further checked whether the matching rate obtained by the correspondence of the vehicle that has passed the upstream / downstream point in the past (for example, once before) is higher than the threshold, and if it is higher, there is no abnormality. to decide. If it is low, it is determined that the sudden event continues.
[0081]
As described above, when the matching rate is continuously higher than the threshold value twice or more, it is determined that there is no abnormality in order to avoid an erroneous determination in the case of accidental matching even if a sudden event occurs.
The threshold value may be determined to an appropriate value by examining the matching rate when an unexpected event actually occurs.
As described above, the matching rate and the determination result of the occurrence of the sudden event are brought into the comprehensive determination described later.
[0082]
Note that other known methods may be employed as a method for calculating the matching rate. For example, a method of taking an image of a vehicle and recognizing the characteristics of the vehicle, such as color and plate number, using image processing may be employed.
3.4 Traffic Flow Abnormality Detection Using Density Waves (Spectrum Change) Appearing at Average Vehicle Speed (Method D)
The determination unit D determines the traffic volume Q (tn) and speed V (at each time t1, t2, t3,..., Tn,. tn), the time occupancy O (tn) is recorded, and their frequency spectrum is calculated. As is well known, this calculation is performed by obtaining an average value of traffic volume Q (tn), speed V (tn) or time occupancy O (tn), obtaining a sample autocovariance shifted by time k, and obtaining a sample spectral distribution pj. If you ask for.
[0083]
Specifically, one of the traffic volume Q (tn), the speed V (tn), or the time occupancy O (tn) is selected and expressed as yn. An average value E (yn) of yn is set to μ.
μ = E (yn) = (1 / N) Σyn (the sum is from n = 1 to N. N is the number of samples)
The sample autocovariance function Ck is
Ck = (1 / N) Σ (yn−μ) (yn−k−μ)
It becomes. However, the sum Σ is from n = k + 1 to N. The sample spectrum pj is
pj = ΣCk exp (-2πikfj)
It becomes. The sum Σ is taken from k = −N + 1 to N−1. In terms of cosine function,
pj = C0 + 2ΣCk cos (2πkfj)
It becomes. The sum Σ is taken from k = 1 to N−1. fj is the frequency,
fj = j / N
j = 0, 1, ..., [N / 2] ([] is a Gaussian symbol)
It is.
[0084]
The determination unit D determines the occurrence of a sudden event based on the sample spectrum distribution.
FIG. 11 is a graph showing fluctuations every minute of the traffic volume Q (tn) calculated on the day when the sudden event occurred. The measurement result of 500m downstream of the sudden event occurrence point is shown.
Q (tn) is stable at around 50 cars / minute before the 8:00 when the sudden event occurred. However, Q (tn) decreases significantly immediately after the time of 8:00 when the sudden event occurs. At 8:20, when the traffic obstruction is removed, Q (tn) rises to about 60 units / minute, and after a while returns to the original steady value of around 50 units / minute.
[0085]
FIG. 12 is a graph that three-dimensionally illustrates the distribution of a sample spectrum (power spectrum) pj obtained for a 30-minute sample (N = 30). The horizontal axis shows frequency and time. The vertical axis is the power spectrum.
FIG. 13 is a graph in which the locus of the peak value of the sample spectrum pj is plotted over time.
The peak value hardly appears before 8:00 when the accident occurred. After 8:00, the peak value rises and keeps the value in the vicinity after 8:20 when the traffic obstacle is removed.
[0086]
FIG. 14 is a graph plotting the peak frequency (unit: Hz) of the sample spectrum. The peak frequency takes a relatively high value before the time of 8:00 when the sudden event occurs, but after the time of 8:00 when the sudden event occurs, the peak frequency drops sharply to 0.05 Hz or less. It remains low even after the traffic obstruction is removed.
FIG. 15 is a flowchart for explaining processing for determining the occurrence of a sudden event performed by the determination unit D. FIG. 15 will be described assuming that the sample spectrum distribution of the traffic volume Q (tk) is obtained.
[0087]
The determination unit D repeats this process every time tn. Describing along FIG. 15, the determination unit D calculates a sample spectrum distribution pj (j is equivalent to a frequency) (step W0). The determination unit D initializes an evaluation value (score) stored in a memory attached to the determination unit D to 0 (step W1), and checks a previous sudden event occurrence determination flag or a traffic state attention determination flag (step S1). W2).
If the previous determination flag is ON, the average value of the previous power spectrum is used as a reference value (step W3), and a constant value α is added to the score (step W4).
[0088]
Next, it is determined whether the peak value of the spectrum (see FIG. 13) is larger than any peak value in the past 30 minutes (step W5). If it is larger, the fixed value β is added to the score (step W6).
Next, it is determined whether or not the change amount (for example, the differential value) of the peak value of the spectrum is larger than the average value for the past 30 minutes (step W7). If it is larger, a constant value γ is added to the score (step W8).
[0089]
Next, it is determined whether or not the peak frequency of the spectrum is smaller than a threshold value (for example, 0.1 Hz) (step W9). If it is smaller, a constant value ε is added to the score (step W10).
The reason for adding to the score is to prevent instantaneous erroneous detection. How to take the constant values α, β, γ, ε is determined automatically after statistically examining the highway to be operated, how accurately sudden events can be actually detected. Can do.
[0090]
Based on the score, the occurrence of a sudden event is determined (step W11), and the determination result is output (step W12). The score and the determination result are used for comprehensive determination.
In the above embodiment, the sample spectrum distribution of the traffic volume Q (tk) is obtained. However, the sample spectrum distribution of the speed V (tk) and the time occupation rate O (tk) can be obtained by the same process. The occurrence of sudden events can be determined. In addition to these, even if the sample spectrum distribution of the vehicle head time interval, the number of vehicles in the road section, the vehicle space occupancy, the space average speed, and the vehicle head distance is obtained, the occurrence of sudden events can be determined by the same process. Yes.
[0091]
In the sample spectrum distribution, in addition to the value of the power spectrum itself and the value of the peak frequency, the same effect can be obtained by using the value of the spectral density and the phase characteristics.
3.5 Conventional traffic flow anomaly detection based on measured values of road traffic volume, vehicle speed, image data, etc. (Method E)
The determination unit E calculates the difference between the measured value of the traffic volume or the vehicle speed at the upstream vehicle sensor 3 and the vehicle sensor 5 and the measured value at the downstream side, and differs by more than this difference threshold value. When an accident occurs, the occurrence of a sudden event in that section is determined, and the result is sent to the overall determination.
[0092]
In addition, the determination unit E performs road image measurement with a camera installed on the road, performs stop / low speed vehicle detection, tracking lane change vehicles, etc. by image processing, and based on this result, an unexpected event such as an accident May be detected, and the result may be sent to comprehensive judgment (see [6] below).
[6] Yamada, Miyao, Sakai, Nishiyama, Ishishita, Negishi “Development of an Abnormal Driving Detection System in Tunnels” Sumitomo Electric No. 145, pp.124-129, September 1994
In addition to the above, a method using a neural network can be employed.
These known methods are collectively referred to as “method E”.
[0093]
4). Comprehensive judgment
The computer 11 is provided with a comprehensive determination unit 23. The comprehensive determination unit 23 calculates and determines the occurrence likelihood (probability) of the sudden event based on the determination results of the plurality of determination algorithms A to E.
4.1 Normalization of evaluation values
In the traffic flow abnormality detection methods A to E described in 3.1 to 3.5, terms are different, but the evaluation value (FIG. 3), correlation (FIG. 4), matching rate (FIG. 10), score ( Fig. 15) Calculates values called differences, sets thresholds in advance, and traffics when these values exceed those thresholds (in the case of correlation, lower than the threshold) An abnormal flow was detected.
[0094]
In the case of “correlation”, in order to be consistent with other methods, it is necessary to take the reciprocal thereof. Further, if the ranges that these values can take are different, comprehensive determination cannot be made, so the ranges that can be taken need to be unified into a certain range. Thereafter, these values are normalized so as to take a range from 0 to 1. Then, the normalized value is referred to as “evaluation value” with a unified name.
Then, the evaluation value calculated by the method A is P _A , The evaluation value calculated by method B is P _B , The evaluation value calculated by Method C is P _C , The evaluation value calculated by method D is P _D , The evaluation value calculated by method E is P _E I will say.
[0095]
These evaluation values P _A ~ P _E The following three determination methods will be described as examples of the comprehensive determination method using the.
In the following description, the known method E is excluded for convenience of explanation, and the description will be based on the four methods A to D. However, the implementation of the present invention is not limited to this embodiment, and any one of the methods A to D can be used. The present invention can also be implemented by replacing method E with method E or an unknown method, or adding method E or unknown method to methods A to D.
[0096]
4.2 Comprehensive judgment 1 (majority method)
This method is a method of performing comprehensive determination based on the number of abnormalities detected by method A to method D.
FIG. 16 is a flowchart for explaining the majority decision method. If all the processes of method A to method D have been completed (step (1)), it is determined whether an abnormality has been detected by three or more methods (step (2)). If an abnormality is detected by the three or more methods, it is determined that the occurrence likelihood of the sudden event is sufficiently high and “the occurrence of the sudden event” (step (6)).
[0097]
If an abnormality is detected only by two or less methods, the evaluation value calculated by each method is added (step (3)). This added value P _A + P _B + P _C + P _D If the threshold value exceeds the threshold value, it is determined that the likelihood of occurrence of the sudden event is moderately high and the “attention state where there is a high possibility of occurrence of the sudden event” (step (7)). If the threshold value is too high, detection failures increase, and if the threshold value is too low, false detections increase. Therefore, the threshold value must be set to an appropriate value. As this threshold value, for example, a value that is 1/2 times the total number of methods (in this case, 4) may be set.
[0098]
If this added value does not exceed the threshold value, the likelihood of occurrence of a sudden event is low, and it is determined that “no sudden event has occurred” (step (5)).
In the above processing, in step (2), it is determined whether or not an abnormality is detected by three or more of the four methods, but the number is not limited to three. It is good. In step (4), a plurality of threshold values may be provided, and a plurality of likelihoods of occurrence of sudden events may be obtained step by step.
[0099]
4.3 Comprehensive judgment 2 (weighting method)
In this method, the evaluation values calculated by the methods A to D are respectively weighted to obtain an average. The weighting coefficients are α, β, γ, and δ, respectively. _total Is displayed.
P _total = (ΑP _A + ΒP _B + ΓP _C + ΔP _D ) / (Α + β + γ + δ)
FIG. 17 is a flowchart for explaining the weighting method. If all the processes of method A to method D have been completed (step (11)), P _total Is calculated. And this P _total Is compared with the detection threshold value (step (2)), and if the detection threshold value is exceeded, the occurrence probability of the sudden event is sufficiently high and it is determined that “the sudden event has occurred” (step (15)). If this detection threshold is too high, detection omissions increase, and if the detection threshold is too low, false detections increase. This detection threshold value may be automatically determined based on the results of the detection omission rate and the false detection rate described later with reference to FIG.
[0100]
If the detection threshold is not exceeded, this P _total Is compared with the attention threshold (step (14)). There is a relationship of attention threshold <detection threshold. If the attention threshold value is exceeded, it is determined that the likelihood of occurrence of a sudden event is moderately high and the “attention state where the possibility of sudden event occurrence is high” (step (16)).
If the caution threshold is not exceeded, the occurrence likelihood of the sudden event is low and it is determined that “no sudden event has occurred” (step (17)).
[0101]
Here, how to determine the weighting factors α, β, γ, and δ will be described. As a premise to make this decision, actual traffic measurement data is collected, and based on the traffic measurement data at the time of the occurrence of the sudden event, abnormalities are detected by each method A to D, and the results such as whether or not the detection is correct are examined. It is necessary to keep.
FIG. 18 is a flowchart for explaining the recording method of the detection rate and the like. First, traffic measurement data is always accumulated (step (21)). When it is known that a sudden event has actually occurred (YES in step (22)), the traffic measurement data before and after the occurrence time is referred to (step (23)), and method A to method D are performed (step ( 24)), it is judged whether the abnormal value of the traffic flow was detected with the evaluation value exceeding the threshold value by each method. The above processing is performed every time an unexpected event occurs.
[0102]
As a result, the probability of correct detection for the total number of sudden events was “correct detection rate”, the probability of failure was not detected for the total number of sudden events was “detection failure rate”, and the total number of detections was incorrect. The probability of detection is referred to as “false detection rate”, and the time from the occurrence of a sudden event to detection is referred to as “detection delay time” (step (25)).
The comprehensive judgment unit 23 classifies and records these values for each method A to method D according to traffic conditions, day of the week, season, weather, and time zone. Table 1 below is an example of evaluating the recorded contents.
[0103]
[Table 1]

[0104]
The comprehensive determination unit 23 determines a weighting factor for the current time.
The weighting factor is a function of traffic volume Q, speed V, occupation rate O, road alignment (curve, zigzag, etc.), day of the week, time zone, degree of traffic jam, past detection results (Table 1), and the like.
FIG. 19 is a flowchart for explaining weight coefficient determination processing. This process is performed in real time. Corrections are made to the initial values of α to δ (for example, all are set to the same value).
[0105]
First, the weight based on the road alignment of the road section to be detected is added (step (31)). For example, if the road alignment is likely to become a bottleneck, the weighting factor α of method A (the number of existing vehicles) and the weighting factor β of method B (lane utilization) are increased.
Next, a weight based on the day of the week is added (step (32)). For example, if the current day is a weekday, the weighting factor α of method A (the number of existing vehicles), the weighting factor β of method B (lane utilization), and the weighting factor γ of method C (matching) are increased. If it is a holiday, the weighting factor δ of method D (spectrum) is increased.
[0106]
Next, a weight based on the time zone is added (step (33)). For example, in the morning, the weighting factor α of method A (the number of existing vehicles) is increased. If it is late at night, the weighting factor β of method B (lane utilization rate) is lowered.
Next, a weight based on past results is added (step (34)). For example, the weight coefficient of a method with a high detection rate in the section is increased.
Then, the current traffic condition (congestion level) is checked (step (35)). If there is no traffic jam (NO in step (36)), the weighting factor β of method B (lane utilization) and the weighting factor δ of method D (spectrum) are increased (step (37)).
[0107]
If there is a traffic jam, it is determined whether the traffic is in the initial stage (step (38)). If the traffic jam is in the initial stage, the weighting factor α of method A (the number of existing vehicles) is increased (step (40)). If the traffic is in the middle to the end, the weighting factor δ of method D (spectrum) is increased (step (39)).
As described above, the weighting factor is automatically determined.
4.4 Comprehensive judgment 3 (detection order method)
This method is a comprehensive determination method focusing on the detection delay time of method A to method D.
[0108]
For simplicity, two methods (method A to method D may be used, but here, method A and method D) will be described as an example.
FIG. 20 is a graph showing the ratio of simultaneous detection of sudden events by Method A and Method D, the ratio of Method A detected earlier than Method D, and the ratio of Method A detected later than Method D based on past results. is there. The ratio of the sudden events detected by the method A and the method D simultaneously is 43%, the ratio of the method A detected earlier than the method D is 34%, and the ratio of the method A detected later than the method D is 23%. The reason for this is that Method A tends to have a high detection rate in the early stage of traffic congestion after the occurrence of the sudden event, whereas Method D tends to have a high detection rate during the traffic jam after the occurrence of the sudden event. It is considered based.
[0109]
FIG. 21A is a graph showing a detection success rate and a failure rate by weighted average when the method A and the method D are detected at the same time. The detection success rate and failure rate are 89% and 11%, for example. FIG. 21B is a graph showing detection success rates and failure rates by weighted average when method A is detected earlier than method D. The detection success rate and failure rate are, for example, 100% and 0%. FIG. 21 (c) is a graph showing detection success rates and failure rates by weighted average when method D is detected earlier than method A. FIG. The detection success rate and failure rate are, for example, 80% and 20%.
[0110]
As described above, since the detection success rate when method A is detected earlier than method D is as high as 100%, according to this detection order method, when method A is actually detected earlier than method D, it is unconditional. It is determined that a sudden event occurred. That is, the likelihood of occurrence of abnormality is set to the maximum value. For the first time in other cases, determination is made by the method of comprehensive determination 1 or comprehensive determination 2.
FIG. 22 is a flowchart for explaining a detection order method determination method. When all the processes of method A to method D are completed (step (51)) and at least the method A and method D detect the occurrence of a sudden event (YES in step (52)), the respective methods A and D The detection time data is referred to (step (53)).
[0111]
It is examined which method is detected earlier (step (54)). At this time, if the detection times of the respective methods are so far apart that they cannot be considered to detect the same sudden event, it is meaningless to make a comprehensive determination. Accordingly, it is preferable that the comprehensive determination 3 is performed only when the detection time difference between the methods is within a certain time or within a certain processing cycle, and otherwise this determination is not performed.
If method A is detected earlier than method B, the likelihood is the highest, and the occurrence of a sudden event is determined (step (55)).
[0112]
If it is detected at the same time or if method B is detected earlier than method A, the occurrence of sudden event is determined using the method of comprehensive determination 1 or comprehensive determination 2 described above (step (56)). When it is determined that a sudden event has occurred (YES in step (57)), it is determined that the likelihood is medium and there is a possibility that a sudden event has occurred (step (58)). When the occurrence of the sudden event is not determined, the likelihood is low and it is determined that the normal state is reached (step (59)).
[0113]
As described above, if the method A is detected earlier than the method B, the comprehensive determination 3 determines the occurrence of a sudden event without performing the comprehensive determination 1 or the comprehensive determination 2 described above, focusing on past trends. There is a feature in doing.
4.5 Identification of sudden event occurrence section
When the occurrence of the sudden event described above is determined in a plurality of sections, the likelihood of determination in each section is compared, and the section with the highest likelihood can be specified as the sudden event occurrence section.
[0114]
When the occurrence of the sudden event and its occurrence section are determined as described above, the traffic management center 10 displays the occurrence of the sudden event such as “Fast Accident / Stop” on the variable display boards 6 and 9, The occurrence of a sudden event is notified to the vehicle through the roadside beacon 7. In addition, the related organization 13 and the broadcasting station 14 are notified through the communication line. Further, in the case of traffic condition attention, the traffic management center 10 displays a message for alerting the driver on the variable display boards 6 and 9 such as “forward attention”, and also to the vehicle through the roadside beacon 7. Notify that it is a driving attention section.
[0115]
If road construction or the like is already planned and an unexpected event is expected, the traffic management center 10 may determine whether an unexpected event has occurred at that time, but based on this determination, the variable display board 6 , 9 are not displayed, and the related organizations 13 and broadcast stations 14 are not displayed.
[0116]
【The invention's effect】
As described above, according to the traffic flow abnormality detection device or method of the present invention, by combining the detection results of the respective methods, the drawbacks of the respective methods can be compensated for, so that the occurrence of sudden events on the road is more accurate. Can be detected well.
[Brief description of the drawings]
FIG. 1 is a schematic diagram showing a traffic flow monitoring system for detecting an abnormality in traffic flow.
FIG. 2 is a functional block diagram of a computer 11 in the traffic management center 10;
FIG. 3 is a flowchart for explaining processing for monitoring occurrence of a sudden event performed by a determination unit A;
FIG. 4 is a flowchart for explaining processing for monitoring occurrence of a sudden event performed by a determination unit B;
FIG. 5 is a flowchart for explaining a vehicle group matching processing method performed by a matching determination unit C;
FIG. 6 is a diagram depicting a two-dimensional grid-like directed graph.
FIG. 7 is a diagram of a two-dimensional grid-like directed graph for showing the result of solving the longest path problem with n = 3 and m = 4.
FIG. 8 is a two-dimensional grid-like directed graph when n = 4 and m = 3 and only the vehicle b4 is not matched.
FIG. 9 is a diagram depicting a two-dimensional grid-like directed graph for explaining an algorithm that does not require an external gap on the time series latest point side.
FIG. 10 is a flowchart for explaining processing for monitoring occurrence of sudden events performed by a determination unit C;
FIG. 11 is a graph showing fluctuations every minute of traffic volume Q (tn) calculated on the day when an accident actually occurred.
FIG. 12 is a graph illustrating a temporal change of a sample spectrum (power spectrum) pj obtained.
FIG. 13 is a graph in which a sample spectrum pj is obtained and the peak values of the spectrum are plotted.
FIG. 14 is a graph plotting a peak frequency of a sample spectrum.
FIG. 15 is a flowchart for explaining processing for monitoring occurrence of sudden events performed by a determination unit D;
FIG. 16 is a flowchart for explaining a majority voting method, which is one of comprehensive determination methods.
FIG. 17 is a flowchart for explaining a weighting method which is one of comprehensive determination methods.
FIG. 18 is a flowchart for explaining a recording method such as a detection rate based on past results;
FIG. 19 is a flowchart for explaining weight coefficient determination processing;
FIG. 20 is a graph showing the ratio of simultaneous detection of sudden events by Method A and Method D, the ratio of detection of Method A earlier than Method D, and the ratio of detection of Method A later than Method D based on past results; is there.
FIG. 21A is a graph showing detection success rates and failure rates based on weighted averages when Method A and Method D are detected simultaneously. (b) is a graph showing a detection success rate and a failure rate by weighted average when method D is detected earlier than method A. (c) is a graph showing detection success rate and failure rate by weighted average when method D is detected earlier than method A.
FIG. 22 is a flowchart for explaining a detection order method determination method;
[Explanation of symbols]
1 highway
2 General road
3 Vehicle detector
4 Primary processing equipment
5 Vehicle detector
6 Variable display board
7 roadside beacons
9 Variable display board
10 Traffic Management Center
11 Computer
13 related organizations
14 Broadcasting stations
21 Input processing section
22, A ~ E judgment part
23 Comprehensive judgment part
25 Output processing section

Claims (9)

An anomaly detecting means for detecting an anomaly of traffic flow in two or more different methods based on traffic measurement data;
Comprehensive determination means for determining the occurrence of traffic flow abnormality by combining the detection results of each method by the abnormality detection means,
The traffic flow abnormality detection device according to claim 1, wherein the comprehensive determination unit performs a comprehensive determination based on a detection time order in which an abnormality is detected in each method being in accordance with an order specific to the combination of the methods. 2. The traffic flow abnormality detection device according to claim 1 , wherein the abnormality detection means detects traffic flow abnormality by two or more methods selected from the following (1) to (4).
(1) Anomaly detection of traffic flow using time fluctuation of the number of vehicles in the road section,
(2) Traffic flow anomaly detection focusing on lane usage rate bias,
(3) Anomaly detection of traffic flow by determining matching of vehicle groups,
(4) Traffic flow anomaly detection using dense waves (spectrum change) appearing in the average speed of the vehicle. The traffic flow abnormality detection device according to claim 1 or 2 , wherein a detection result in each method is obtained based on traffic measurement data before and after the occurrence of a sudden event and is accumulated as actual data. . 4. The traffic flow abnormality detection device according to claim 3 , wherein the actual data includes one or more data among a positive detection rate, a detection omission rate, a false detection rate, and a detection delay time. Abnormality detection apparatus according to claim 1 or claim 2, wherein the traffic flow, characterized by further comprising information providing means for outputting a result of determination by the overall judging means as the abnormality information. The comprehensive determination unit performs stepwise determination according to the magnitude of the value that is the basis of the determination, and the information providing unit determines the content of the abnormality information based on the result of stepwise determination by the comprehensive determination unit The traffic flow abnormality detection device according to claim 5, wherein the traffic flow abnormality detection device is changed. The said comprehensive determination means, when the abnormality of traffic flow is detected in a plurality of road sections, specifies the section where the abnormality has occurred according to the magnitude of the value which is the basis of the determination. The traffic flow abnormality detection device according to claim 6 .   Based on traffic measurement data, traffic flow anomalies are detected in two or more different methods, and the detection time ranks detected in each method are in accordance with the order specific to the combination of the methods, and traffic flow anomalies are detected. An abnormality detection method for traffic flow, characterized by determining occurrence. 9. The traffic flow abnormality detection method according to claim 8, wherein the two or more different systems include two or more systems selected from the following (1) to (4).
(1) Anomaly detection of traffic flow using time fluctuation of the number of vehicles in the road section,
(2) Traffic flow anomaly detection focusing on lane usage rate bias,
(3) Anomaly detection of traffic flow by determining matching of vehicle groups,
(4) Traffic flow anomaly detection using dense waves (spectrum change) appearing in the average speed of the vehicle.

Images