辨識與計數物件流量之嵌入式深度學習系統

本篇論文中，通過嵌入式系統(embedded system)實現深度學習(deep learning)，設計辨識與計數的系統，提出了追蹤車流以及人流的計數方案。本論文分成三部分，第一部分，通過神經運算棒加強的深度學習做目標辨識，目標物件為道路常見的交通工具，車子、以及行人，第二部分，建構了目標追蹤法，依據在視訊流的連續幀中，比較已知目標和新出現目標之間的歐氏距離，持續追蹤目標到檢測區或消失。第三部分，在螢幕上設置感興趣區，當目標的辨識及追蹤完成後，系統會根據閥值進行資料處理。本篇論文的研究貢獻如下: 1. 耗費低成本且易部屬多數做深度學習運算皆仰賴運算能力較高的 CPU 以及顯示卡，嵌入式系統售價低廉且體積小，使用免費的 Python 進行程式編譯。 2. 節省人力取代人力在街頭使用計數器。 3. 有效辨別目標目前道路計數方法是使用車輛通過路面下安裝的感測器，但此方法無法辨識通過目標的種類。 4. 流量數據取得流量數據，整合大數據，為智慧化城市和物聯網發展作貢獻。關鍵字:目標物件辨識、目標物件追蹤、深度學習、質心追蹤法、嵌入式裝置、流量分析、神經運算棒、YOLOV3-Tiny、Open VINO。

關鍵字

目標物件辨識；目標物件追蹤；深度學習；質心追蹤法；嵌入式裝置；流量分析；神經運算棒； YOLOV3-Tiny ； Open VINO

並列摘要

In this thesis, through the implementation of deep learning in the embedded system, a system for identification and counting is designed, and a counting scheme for tracking the flow of vehicles and people is proposed. The study is constituted by three parts. In the first part, target recognition through deep learning enhanced implemented by neural compute stick, targeted objects are pedestrians and common vehicles on the road, such as cars. In the second part, A target tracking method is architected, which the method compares the Euclidean distance between a known target and a new-emerging target in successive frames of a video stream, and continuously tracks the targets in the whole video screen until they disappeared. In the third part, based on the function library of OpenSource Computer Vision (OpenCV), a designated inspecting area is set up in the video screen. Once the recognizing and tracking of the target are completed, the system processes the data of the targets according to the threshold from its region of interest. The contribution from the study is as below: 1. Low cost and easy to deploy Majority of the behaviors of deep learning relies on using high-computing-power CPUs and graphics cards. The embedded systems are cheap and in small size. The compilation for the program can be implemented with free programming language, Python! 2. Reduce labor cost With machine and hardware and software to replace counting by labor resource. 3. Effectively recognize the targeted objects At present, the counting method is by detecting the passing objects with sensors installed beneath the road surface, but the method cannot identify the class of the objects. 4. Traffic data can be collected for further application Obtain traffic data then integrate into big data, the results from analyzing can be further applied on enhancing the development of smart cities and the Internet of Things. Keywords: Object detection、Object tracking、Deep-learning、YOLOv3-Tiny、Centroid tracking、Embedded device、Open VINO、Neural Compute Stick、Traffic Analysis.