以支持向量機為基礎即時自我學習物體偵測系統之演算法與架構設計

Abstract

近幾年來，視覺辨識技術參予了許多影片及圖片分析等智慧型應用，而像是機器控制、智慧型環境監控、人類動作辨識都因為電腦視覺的演進而變的可行，其中機器學習扮演了重要的角色，機器學習已經變成電腦視覺應用中的基本核心模組，而我們相信如果我們能對機器學習的領域做任何可能的貢獻，我們就能將電腦視覺的進展推向新的世代。 雖然我們的主要目標是要在機器學習做出突破，但我們暫時將目標專注在物體辨識，我們相信藉由視覺相關的應用所呈現的結果較能展現機器學習的進步所帶來的效能，而現在的物體偵測依然有許多的限制，我們需要許多的訓練資料來達到好的物體偵測模組，但是因為人為標誌的資料太過昂貴而且花時間，導致藥準備所有我們希望能夠偵測的物體資料是不可能的。因此我們需要找出一種方法可以讓物體偵測系統有效的運用少量的資料進而線上學習，更好的狀況是讓系統可以自己學習新的資料並更新現有的模組。 在這篇論文中，我們呈現了數種機器學習的演算法架構，同時我們也對他們做了全方位的比較。我們所開發的系統可以偵測追蹤物體，並且自己更新核心模組，這部份最主要的貢獻在於我們藉由這個物體辨識系統證明自動學習的可能性，並且我們將該系統原本的核心學習演算法改成較有擴展性並適合即時應用的自動向量機(SVM)。 最後，我們呈現了一個全新並適合硬體實現且基於SVM的學習概念，並大大的改進了訓練以及預測的效能，這兩個概念分別是漸進式壓縮及叢林式投票，他們同時改進了記憶體使用量、線上運算的時間。 整體的來說，我們發展出了一個可以自動更新學習核心的物體偵測系統，同時我們針對SVM做出了重大的突破。

In recent years, visual recognition techniques are involved in efficiently analyzing video or image data for most of the intelligent applications. And the new applications like robot control, intelligent environmental surveillance, human action analysis, and in-home robotics are all becoming possible due to the advancement of computer vision developments, and machine learning plays a particular essential role in all those tasks. Machine learning becomes a basic core module for computer vision applications. The application in computer vision can be pushed into a new era if we could make any possible improvement in machine learning research area. Though the main target is to make some breakthrough in machine learning, we focus on object detection as our application. We believe that the improvement of performance in machine learning can be observed easier if the result is present with vision related application. There are still limitations in current object detection methods. Huge amount of training data is needed for learning to achieve a good performance detection model. However, it is almost impossible to prepare all of desired datasets for all types of objects which we want to detect since the human labelled data is too expensive and the types of real-world object is too variety to be fully covered. Therefore, we need to figure out a way to make the robot or an object detection system feasible to learn a new knowledge with few training dataset on-line. Moreover, it will be better if the system can learning the new data type or update the existing model automatically by itself. In this thesis, the algorithm of several important learning algorithm and architecture are presented. And the comprehensive comparison between all of them are provided. Also, we developed a system which is capable of detecting and recognizing the object, moreover updating the core model by itself. The most important contribution of this part is to verify the possibility of automatic learning and we further substitute the core learning model with modified SVM which we proved to be more extensible and to be feasible in real-time application. Finally, a novel hardware-oriented learning concept based on SVM is proposed to improve the performance of both training stage and predicting stage. The two concepts are Adaptive Condense and Forest-like Voting which improve the performance in memory usage, computation time in training stage and are proved to be more suitable for on-line learning application comparing to original SVM. Totally speaking, we present an object detection system able to update the model by itself and we further make a huge breakthrough in Support Vector Machine which is one of the most prevailing learning algorithm.