日照条件に頑健な交通流動画像からの四輪車の自動抽出手法の開発

Abstract

本論文では，車両や建物の影といった局所的な日照変化にも頑健で，昼夜間を問わず適用可能な交通流からの四輪車の位置と軌跡の自動抽出手法を提案する．提案手法は時間差分画像を一定時間内で累積した累積時間差分画像と呼ぶ画像を作成し，閾値以上の輝度値の変化回数を計数して影領域の大半を除去する．この際に車両の一部が欠損するが，累積時間差分画像を更に重ね合わせて車両領域の欠損を補い車両軌跡を決定する．この影を含まない軌跡領域とある時刻における影を含む車両領域との積を取ることで，影なしの車両領域を決定する．最後に上方から見下ろす鳥瞰画像に変換し，車両領域の位置の差分を計算することで車両速度を推定する．歩道橋で撮影した交通流画像に提案手法を適用した結果，昼夜間を問わず高精度かつ安定的に機能することが判明した．

This paper presents a novel method that automatically extracts regions and tracks of vehicles from traffic video images observed even when local sunlight condition varies or at nighttime. The proposed method first generates “accumulated temporal subtraction images” by accumulating temporal subtraction images within a certain period, and removes shadows by counting the number of brightness change in the images. This process removes parts of vehicles, but the lacking parts are compensated by accumulating the accumulated temporal subtraction images after removing shadow. As a result, the vehicle's track is determined. Then, the vehicle region at a certain time is delineated by multiplying the track without shadow and a tentative vehicle region with shadow. Finally, the image is converted into bird-eye view images, and the velocity of a vehicle is estimated by subtracting vehicle's positions shown in two bird-eye view images. We applied the method to traffic video images observed in Kyoto, and it was found that F-value was 0.98, and accordingly the proposed method performs very accurately.