| 摘要: | 深度學習 (deep learning) 或稱為卷積神經網路 (convolutional neural network, CNN) 最近幾年因 顯著的效果改進而極度的熱門。雖然卷積神經網路的基礎是來自90年代的類神經網路 (artificial neural network)，但因為快速硬體的推波助瀾，並且加入卷積運算，藉由學習取得最好的特徵代替傳統神經 網路的人工特徵輸入，使得神經網路的分類與辨識應用變得又快又準確。2016年我們做了卷積神經網 路的人臉辨識；相較於傳統的主成份分析 (principal component analysis, PCA) 之特徵臉 (Eigenface)、 線性分類 (linear discrimination analysis, LDA) 的費雪臉 (Fisherface)、支援向量機 (support vector machine, SVM)、及適應性提昇分類器 (adaptive boost, AdaBoost)，卷積神經網路都有顯著的效果改進， 辨識率約從 70 % 提昇到 99 %。 從2001年起到現在，我們就一直很積極的從事先進駕駛輔助系統 (ADAS) 之視覺偵測與辨識技術 的實務研究，其中我們完成 12 項軟體技術：1.車道偏離警示 (LDW)、2.前車碰撞警示 (FCW)、3.盲 點偵測 (BSD)、4.行人碰撞警示 (PCW)、5.交通標誌/號誌偵測與辨識 (TSSR)、6.全周俯瞰監視系統 (STM)、7.廣域全周俯瞰監視與偵測 (WSTD)、8.影像式停車導引 (IPG)、 9.倒車碰撞警示 (RCW)、10. 影像式主動跟車偵測 (ISG)、11.自動跟隨巡航 (AFN)、及 12.昏睡偵測 (DD)。其中部份技術已技術轉 移給廠商，且部份技術也已變成商用產品。 在上述先進駕駛輔助系統 (ADAS) 的研究項中，我們有碰到一些偵測率一直無法改善的問題；例 如，在不良天候的前車偵測、行人偵測的穩定度與可靠度、倒車障礙物的偵測、自動跟隨追蹤的穩定 度、.. 等。因此才促成我們提出這一個三年期的研究計畫，擬以深度學習技術提升部份傳統先進駕駛 輔助系統的偵測與辨識率。在這個計畫中，每一年都有一個卷積神經網路的技術發展項目及一個先進 駕駛輔助系統的應用項目。第一年我們將發展適合台灣IC設計製造的小網路系統架構，解決擬合不足 (underfitting) 與擬合過度 (overfitting) 的問題，並且應用於前車碰撞警示 (FCW) 系統上。第二年發 展快速物件偵測的卷積神經網路系統，在不需要使用滑動視窗掃描整張影像下，同時提高物件偵測效 果 (effect) 與速度 (performance)，並且應用於行人碰撞警示 (PCW) 系統與倒車碰撞警示 (RCW) 系 統上。第三年發展小樣本的物件偵測與追蹤之卷積神經網路系統，使得卷積網路可在不同環境與場景 下快速轉移應用，並且應用於自動跟隨巡航 (AFN) 系統的前導者偵測與追蹤上。本研究亦將卷積神 經網路系統實現在行動嵌入式平行處理器上。 本研究是建立在過去的基礎及成果上，針對特定議題，發展深度卷積神經網路技術解決過去難以 顯著突破的偵測與辨識問題。計畫主持人已有二十多年電腦視覺的研究經歷，且已有十多年電腦視覺 技術應用在駕駛輔助系統上的經驗；更在 2008 - 2010 年間合聘到工研院機械所智慧車輛組協助先進 安全車輛之視覺偵測技術的發展，並且獨自與共同獲得或申請多個電腦視覺輔助道路安全駕駛之中華 民國、美國、及中國發明專利，其部份成果具備高度實用性，已被數個單位及公司選用於發展產品。 在 2015 年，我們與台積電合作研究深度學習在顯微影像 (scanning electron microscope, SEM) 的瑕疵 檢測應用，在 2016 年也與國安單位合作深度學習在人臉辨識上的應用，因此我們有信心及能力完成本計畫的執行。 ;Recently, deep learning (DL) or convolutional neural network (CNN) has become the hottest topic in information technology due the big success in the applications of artificial intelligence and pattern recognition. Basically, CNN is just revolved from the 90’s traditional artificial neural network; however, it adds the convolutional layers to acquire best features instead of manual features to make the classification and recognition dramatically improved. In 2016, we processed a face recognition study based on a CNN. Comparing with the traditional approaches for human face recognition: Eigenface based on the principal component analysis (PCA), Fisherface based on the linear discrimination analysis (LDA), support vector machine (SVM), and adaptive boosting (AdaBoost), the CNN obtained obvious improvement; the recognition rate was upgraded from about 70 % to 99 %. From 2001 to now, we have studied the visual detection techniques for the advanced driver assistance system (ADAS). In this period, we have completed 12 software techniques: 1.lane departure warning (LDW), 2.forward collision warning (FCW), 3.blind spot detection (BSD), 4.pedestrian collision warning (PCW), 5.traffic sign and signal recognition (TSSR), 6.surrounding top-view monitor (STM), 7.wide-scopic surrounding top-view monitor and detection (WSTD), 8.image-based parking guiding (IPG), 9.rear collision warning (RCW), 10.image-based stopping and go (ISG), 11.automatic following navigation (AFN), and 12.drowsiness detection (DD). Partial programs had been transformed to companies and partial systems had become commercial products. In the above ADAS studies, we encountered several tricky problems, such as the bad detection rate of front vehicle detection in the bad weather conditions, the unstable pedestrian detection, the unreliable obstacle detection in the rear monitor, the unpredicted results in the automatic following navigation, etc. Thus, we propose a three-year research project, intending using the techniques of CNN to improve the detection and recognition rates of our ADAS studies. In this project, we have a technique development and an ADAS application in each year. In the first year, we will: i.develop a small CNN structure to match the ability of the IC design in Taiwan, ii.solve the problems of underfitting and overfitting, and iii.apply to the FCW system. In the second year, we want to: i.develop a CNN system for high-speed object detection without scanning the whole image, and ii.apply to the pedestrian detection in the PCW system and the obstacle detection in the RCW system. In the third year, we intend to i.develop a small-sample CNN system for target detection and tracking, such that the CNN system can be easily transformed and applied among different situations, and ii.apply to the guider detection and tracking in the AFN system. Moreover, we will also implement the proposed CNN systems on the mobile embedded parallel devices. This study is based on our fruitful previous results, focus on the fixed topics to develop special CNN systems to solve the tricky problems on visual detection and recognition. The principal investigator of this project is an original researcher on computer vision; he has studied computer vision techniques more than thirty years; moreover, he has the experience of computer vision applied on ADAS more than fifteen years. He has gotten 10 US, Taiwan, or China patents in these few years. Partial techniques are practiced and have been employed by several companies. In 2015, we collaborated with TSMC (Taiwan Semiconductor Manufacturing Company) to study the CNN technique on the defect detection of scanning electron microscope (SEM) images. In 2016, we also studied CNN technique on human face recognition with national security unit; thus we have ability to complete the execution of this research project. |
