劉強 謝謙 方璽 李波 蔣瓊 解孝民



【摘要】為實現更快速、準確的疲勞預警,提出了一種基于并行短時面部特征的駕駛人疲勞檢測方法。基于加入了 ? MicroNet模塊、CA注意力機制、Wise-IoU損失函數的YOLOv7-MCW目標檢測網絡提取駕駛人面部的短時面部特征,再使用并行Informer時序預測網絡整合YOLOv7-MCW目標檢測網絡得到的面部時空信息,對駕駛人疲勞狀態進行檢測與預警。結果表明:在領域內公開數據集UTA-RLDD和NTHU-DDD上,YOLOv7-MCW-Informer模型的準確率分別為97.50%和94.48%,單幀檢測時間降低至28 ms,證明該模型具有良好的實時疲勞檢測性能。
主題詞:智能交通 疲勞檢測 目標檢測 注意力機制 時序預測
中圖分類號:U492.8+4 ? 文獻標志碼:A ? DOI: 10.19620/j.cnki.1000-3703.20230617
Research on Driver Fatigue Detection Method Based on Parallel Short-Term Facial Features
Liu Qiang1, Xie Qian1, Fang Xi2, Li Bo3, Xie Xiaomin4
(1. School of Intelligent Systems Engineering, Sun Yat-sen University, Shenzhen 518107; 2. Development & Research Center of State Post Bureau, Beijing 100868; 3. Automobile Engineering Research Institute of Guangzhou Automobile Group Co., Ltd., Guangzhou 511434; 4. Guangdong Marshell Electric Technology Co., Ltd., Zhaoqing 523268)
【Abstract】A driver fatigue detection method based on parallel short-term facial features is proposed to achieve faster and more accurate fatigue warning. The method utilizes the YOLOv7-MCW object detection network, which incorporates the MicroNet module, CA attention mechanism, and Wise-IoU loss function, to extract short-term facial features of the drivers face. The parallel Informer temporal prediction network is then used to integrate the spatiotemporal information obtained from the YOLOv7-MCW object detection network, enabling the detection and warning of driver fatigue. The results demonstrate that the YOLOv7-MCW-Informer model achieves accuracy rates of 97.50% and 94.48% on the publicly available datasets UTA-RLDD and NTHU-DDD, respectively, with a single-frame detection time reduced to 28 ms, proving the excellent real-time fatigue detection performance of the model.
Key words: Intelligent transportation, Fatigue detection, Object detection, Attention mechanism, Time series prediction
【引用格式】 劉強, 謝謙, 方璽, 等. 基于并行短時面部特征的駕駛人疲勞檢測方法研究[J]. 汽車技術, 2024(5): 15-21.
LIU Q, XIE Q, FANG X, et al. Research on Driver Fatigue Detection Method Based on Parallel Short-Term Facial Features[J]. Automobile Technology, 2024(5): 15-21.
1 前言
基于駕駛人面部特征的疲勞檢測方法因具有快速、準確的優點被廣泛用于交通安全研究。駕駛人的面部特征主要包括單位時間內閉眼百分比(Percentage of Eyelid Closure Over Time,PERCLOS)[1]、眨眼頻率、視線方向、單位時間內張口百分比(Percentage of Mouth Open Over the Pupil over Time,POM)[2]、哈欠頻率、點頭次數和頭部偏轉角等。在較短的單位時間(一般為1 min)內具有較為明顯的變化規律的面部特征,本文稱為短時面部特征,如PERCLOS、POM等。
國內外圍繞基于短時面部特征的駕駛人疲勞檢測展開了相關研究。Bai等[3]提出使用雙流時空圖卷積網絡檢測駕駛人疲勞,采用面部標志檢測法從實時視頻中提取駕駛人面部標志,然后通過雙流時空圖卷積網絡得到駕駛人疲勞檢測結果,試驗表明,該方法顯著提高了疲勞檢測性能,準確率高達92.70%,但該模型結構較為復雜,檢測時間較長,實時性不足。……