| 摘要: | 駕駛人異常行為偵測是近年來常被關注的熱門議題，透過系統偵測駕駛人的面部姿態、其他肢體以判斷當前駕駛人的狀態，若出現分心、疲勞駕駛等異常行為，系統會發出警告聲提醒駕駛人，以減少車禍傷亡的潛在風險。由於在真實場景中的車況變化多端，因此偵測系統必需同時具備「高偵測準確度」以及「能夠及時偵測」這兩項能力。
近年來深度學習在電腦視覺領域中取得相當的成功，至今已有許多文獻利用電腦視覺技術方法解決駕駛人異常行為偵測的議題，然而現有的文獻方法中，在「精度」以及「速度」兩方面無法同時取得良好的結果。本論文我們基於動作辨識文獻[15]中結合2D與3D卷積神經網路的方法，此架構擁有良好的靜態及動態特徵擷取能力，我們以此作為共享網路，同時學習昏睡、打盹、打哈欠、以及抽菸類別，並搭配其他正則化技巧像是Batch Normalization、預訓練、資料擴增等，以進一步提升效能。此外，我們也基於文獻[15]設計了一套在線視頻動作預測演算法，此演算法除了可以有效率地處理在線視頻預測問題，也能提取更長的時序結構以提升動作偵測的準確度。
在實驗中我們探討了預訓練、資料擴增、網路複雜度、採樣幀數大小、以及不同網路架構方法對模型的效能影響，我們證實了使用預訓練、資料擴增能有效幫助提升模型的精度，且若使用更多種資料擴增組合，精度也能更進一步提升。在採樣幀數與不同網路架構方法的實驗中，我們也證實了本論文架構能達到即時的預測速度，且在精度與模型大小方面皆比2D CNN方法及3D CNN方法好。
;Driver abnormal behavior detection is a hot issue that has been paid attention in recent years, the system detects the driver′s facial posture and other limbs to determine the current driver′s state. If abnormal behaviors such as distraction and drowsy driving have been detected, the system will make a warning sound to remind the driver to reduce the potential risks of car accident casualties. Since the vehicle conditions in real scenes change a lot, the detection system must provide both "high detection accuracy" and "capable of real time detection" abilities.
In recent years, deep learning has achieved considerable success in computer vision field. So far, there have been many papers using computer vision techniques to solve driver abnormal behavior detection issue. However, in the existing literature methods, both "accuracy" and "speed" cannot achieve good results at the same time. In this paper, we based on the method that combining 2D and 3D convolutional neural network from action recognition literature[15], this network can extract both static and dynamic features well. We regard this network as a shared network, and learn drowsiness, nodding, yawning and smoking classes simultaneously, as well as other regularization techniques like batch normalization, pre-training, data augmentations, etc. to further improve performance. In addition, we also design online video action prediction algorithm based on [15], this algorithm not only can handle online prediction problems efficiently, but also can extract longer-range temporal structure, further improving action detection accuracy.
In the experiment part, we explore the effects of pre-training, data augmentations, network complexity, the size of sampling frames, and different network architecture methods in terms of model performance. We confirm that we can gain significant accuracy improvements by applying pre-training and data augmentations, and if more types of data augmentations are used, the accuracy can be further improved again. Furthermore, according to the experiments of sampling frames and different network architecture methods, we also confirm that our model can not only achieve real time inference speed but also outperform 2D CNN and 3D CNN methods in terms of both accuracy and model size. |
