@article{oai:ipsj.ixsq.nii.ac.jp:00240016, author = {中嶋, 健翔 and 西尾, 信彦 and Kensho, Nakajima and Nobuhiko, Nishio}, issue = {10}, journal = {情報処理学会論文誌}, month = {Oct}, note = {自動運転において,Light detection and ranging(LiDAR)は周囲環境を認識するために必要不可欠なセンサである.しかし降雪下では,降雪にLiDARのレーザが反射することで生じるノイズが,自動運転システムの下流タスクである物体検出の精度に影響を及ぼすことが知られている.この問題を解決するために,降雪ノイズの除去に関する研究が行われているが,既存手法の降雪ノイズの除去精度や処理時間は自動運転システムに導入するには不十分である.そこで,本研究は各点において近傍点を探索する既存手法とは異なり,降雪ノイズの特徴を分析した結果に基づいて,信号強度と距離,高さにおいて閾値を設けることで降雪ノイズを除去する新たな手法を提案する.評価では,提案手法の有効性を検証するために異なる環境で取得された複数のデータセットを用いた.Winter Adverse Driving dataSet(WADS)を用いて,除去精度と処理時間について評価した.除去精度については降雪ノイズ検出のRecallとPrecision,F値を計算した.評価の結果,既存手法と同等のRecallであると同時に,Precisionが約13.7%,F値が約7.4%向上したことを確認した.また,提案手法は既存手法の処理時間を約93.7%短縮し,大幅な処理の高速化を実現した.それに加え,Canadian Adverse Driving Conditions(CADC)データセットを用いて降雪下での物体検出について検証した結果,提案手法によって検出漏れが少なくなり,本手法が降雪下での物体検出精度の向上に有用であることを示した., This study introduces a novel method for removing snowfall noise in autonomous driving systems. Light detection and ranging (LiDAR) is crucial for environment recognition in autonomous vehicles, but noise generated by the reflection of LiDAR laser beams on falling snow can impact the accuracy of downstream, such as object detection. Existing snowfall noise removal techniques, although useful, lack the necessary accuracy and processing speed for effective integration into autonomous systems. Our method sets thresholds based on intensity, distance, and height parameters, diverging from traditional methods that rely on searching for k-nearest neighbors at each point. This new approach was developed after thorough analysis of snowfall noise characteristics. To test its efficacy, we employed Winter Adverse Driving DataSet (WADS). Our evaluations focused on noise removal accuracy and processing speed. The results were promising. Our method not only matched the recall levels of existing techniques but also improved precision by approximately 13.7% and F-measure by about 7.4%. Most notably, it significantly reduced processing time by approximately 93.7%, marking a major advancement in speed. Further validation using the Canadian Adverse Driving Conditions (CADC) dataset confirmed the method's effectiveness in enhancing object detection under snowfall conditions. This advancement holds significant potential for improving the reliability and performance of autonomous driving systems in adverse weather.}, pages = {1523--1532}, title = {自動運転における降雪ノイズに頑強な物体検出}, volume = {65}, year = {2024} }