應用於先進駕駛輔助系統之多重物件偵測深度學習技術與其嵌入式平台實現

Abstract

本論文提出了改善使用深度學習方法的行人、汽車偵測技術,在有限的精準度損 失之下能夠提升原本演算法的執行速度,本論文所提方法朝著三個方向進行最佳化,第 一為學習參數的設定、第二為深度學習架構的精簡,因為現有著名的深度學習架構皆 是用來分類數百、數千個物件,本方法針對兩種物件做分類,因此其架構不需要太複 雜,第三為樣本的篩選。透過這一系列的優化方法,我們已成功訓練出一個在有限精 準度損失下、可提升執行速度並於嵌入式系統上實現之深度學習網路。 本系統以個人電腦驗證所提出的設計方法之正確後,並且使用開源的資料庫做驗 證,可以達到 90%的精準度。本系統已同時實作於個人電腦(搭配 GPU)與嵌入式系統 上,在裝置 intel i7@3.5GHz CPU 和 NVIDIA TITANX 的個人電腦上,本系統針對 720x480 解析度影像輸入下其執行效能可達每秒 25 張(25fps),而在 NVIDIA JETSON TX1 配備 ARM Cortex®-A57 和 NVIDIA MaxwellTM GPU with 256 NVIDIA® CUDA® Cores 的嵌入式平台上,本系統針對 720x480 解析度影像輸入下其執行效能可達每秒 5 張(5fps)。

This thesis proposes an optimized pedestrian and vehicle detection method based on deep learning technique. We optimize the convolutional neural network architecture by three mainly methods. The first one is the choice of the learning policy. The second one is to simplify the convolutional neural network architecture. The last one is careful choice of training samples. With limited loss of accuracy, we can greatly speed up the original deep learning method coming from CAFFE. The proposed system is developed on PCs and implemented on the platforms of both the PC and embedded systems. We can achieve around 90% accuracy when it is tested on an opensource dataset. On PCs with Intel i7@3.5GHz CPU, the proposed design can reach the performance about 720x480 video at 25 frames per second. On the NVIDIA JETSON TX1 embedded system equipped with ARM Cortex®A57 and NVIDIA MaxwellTM GPU with 256 NVIDIA® CUDA® Cores, the proposed design can reach the performance about 720x480 video at 5 frames per second.