使用多線索交互和部位感知監督增強之輕量級深度卷積網路應用於面部資訊偵測

Abstract

近年來，智慧車輛相關技術正迅速發展，其中設計能在車輛上運行的駕駛者監控系統成為一項重要課題。這類系統透過分析駕駛者的面部資訊來判斷其當前的注意力狀態，幫助評估駕駛者控制車輛的能力。 在本論文中，我們提出了一種輕量級的多任務深度學習模型，該模型利用面部圖像同時檢測頭部姿態、注視方向和眼睛狀態，專為整合到駕駛者監控系統而設計，旨在判斷駕駛者是否專注於道路。為了使模型能夠部署於車輛上的嵌入式裝置，我們採用多任務學習並設計了合適的任務分支，創建一個高效且輕量級的系統。每個任務由專門的分支支持，以確保提取必要的特徵，並通過部位感知監督增強對相關面部區域的關注。此外，我們引入線索特徵來編碼各部位特定的隱式知識，提供更佳的任務特定特徵初始化並提升整體模型性能。我們在AFLW2000、BIWI、Gaze360 和 CEW 數據集上對我們的模型進行評估，結果顯示我們的方法在減少參數成本的同時，達到了具競爭力的性能，並在相似的輕量級條件下表現更佳，展示了我們的方法在實際應用中的有效性。

In recent years, technologies related to intelligent vehicles have seen rapid devel-opment, making the design of driver monitoring systems that can operate on vehicles acritical area of focus. These systems analyze facial information of drivers to determinetheir current attention status, helping assess the driver’s ability to control the vehicle. In this thesis, we propose a lightweight, multi-task deep learning model designed to simultaneously detect head pose, gaze direction, and eye state from facial images. This model is specifically designed for integration into driver monitoring systems, aiming to determine whether the driver is focused on the road. To facilitate deployment on embedded vehicle devices, our approach leverages multi-task learning with carefully designed task branches to create an efficient and lightweight system. Each task is supported by dedicated branches to ensure the extraction of necessary features, with part-aware supervision enhancing the focus on relevant facial regions. Furthermore, we introduce clue features to encode part-specific implicit knowledge, providing improved initialization for task-specific features and enhancing overall model performance.We evaluate our model using the AFLW2000, BIWI, Gaze360, and CEW datasets, demonstrating that our method achieves competitive performance with reduced parameter costs. Our results show superior performance under lightweight conditions compared to existing methods, demonstrating the effectiveness of our approach in real-world applications.