以混合深度學習方法加速嵌入式系統之物件偵測

Abstract

現今有許多電腦視覺在嵌入式系統上的應用，然而電腦視覺的應用需要大量的計算及功耗，這並不樂見於嵌入式系統。一般而言，使用者在不同應用程式、不同的情境下對於效能及功耗有不同的需求，因此我們想要量測不同演算法在不同平台下效能、功耗和準確度的特性，並透過我們的結果提供使用者一些平台和演算法選擇上的建議。 本研究建立一個以耗能(Power)、準確度(Accuracy)以及幀率(FPS)所組成的三維圖，提供使用者在不同情境下的選擇，並進一步從微觀的角度，利用影像前後幀(frame)之間差異不大的特性，我們挑選出一個辨識準確度較低但辨識時間較短的演算法以及一個辨識準確度較高但辨識時間較長的演算法，讓影像的每幀穿插使用兩演算法，並且以準確度高的結果對準確度低的結果進行修正，進而達到準確度接近準確度較高的演算法但其執行時間大幅縮短的效果。

Nowadays, there are a plenty of the computer vision applications on embedded system. However, it is not good to run computer vision applications on embedded systems which will take much computation and power consumption. This study aims to measure the power consumption, the accuracy and the performance of different algorithm running on different platform to give the suggestion of choosing algorithm on the platform so to meet the request of the performance and power consumption for different applications in different situation. In this thesis, we build a 3D plot which composed by power, accuracy and FPS. Further, from microscopic point of view, we choose one higher accuracy but slower detection algorithm and one lower accuracy but faster detection algorithm. Since video has the characteristics that the frame in video will be similar to its next frame, we run these two detection algorithm on the video frames alternatively and use the result produced by higher accuracy algorithm to correct the result produced by lower accuracy algorithm so that we can achieve the result that the accuracy is close to the higher accuracy algorithm but its execution time significantly shortened.