基於分塊模型與影像路徑規劃之行人檢測與追蹤系統

Pedestrian detection and tracking system using deformable part model and visual-based Trajectory

曹維雨(Wei-yu Cao)

指導教授 ： 連豐力

國立臺灣大學/電機資訊學院/電機工程學研究所/碩士(2017年)

https://doi.org/10.6342/NTU201700140

摘要

近年來,無人機由於其結構簡單,飛行穩定等特點,得到了廣泛的關注。在 民用領域中,人們可以利用它在比較艱難的環境中拍攝影片。然而無人機通常是 由人來控制的,如果一個人獨自行動,並想記錄下這段時光,就會變得很困難。 這篇論文將致力於解決這個問題。
首先要解決的問題是,如何有效的檢測到所有的行人。這篇論文提出了一種 基於分塊模型的檢測方法。在影像中,系統會檢測算法中設定的人體的不同部分, 然後根據人體比例對他們做匹配,來檢測它們是否屬於同一個人。利用這個原理, 可以使原本的檢測方法更加準確,結果也會更加穩定。在訓練分類器的環節中, 負樣本中檢測到的正樣本會用來當作新一輪的副樣本,這個作法也會讓單次檢測 的準確性大幅提升。實驗證明這個行人檢測系統運行的結果良好。
在正確檢測到影像中所有的行人後,接下來便是追蹤環節。這篇論文提出了 一個基於影像的追蹤方法,即路徑規劃方法。系統首先會用一種增強型的匹配算法來匹配特徵點,在原有歐式距離匹配的基礎上,提升了匹配特徵點的數量,同 時移除了明顯錯誤的匹配。然後利用上一張影像的輸出當做關鍵區域,去找下一 張影像中的特定目標,同時加入迴授算法保證追蹤的準確性。在路徑規劃方面, 一些人為定義的理想位置將會在螢幕上作為四旋翼飛行的參考。如果螢幕上的目 標不符合理想位置,飛行器就會進行相應的移動,以保證目標在螢幕上處於理想 的位置。同時,在飛行過程中,系統會即時與預先拍攝好的三個方向的特徵點做 匹配,然後計算目標的正面是否有比較多的百分比在螢幕上。如果系統檢測到目 標有轉動特定角度,四旋翼會立刻旋轉,然後進行一系列的動作,飛回到目標正 前方的位置,以保證目標行人的正面呈現在螢幕上。

關鍵字

無人機 ； 行人檢測 ； 目標追蹤 ； 路徑規劃 ； 特徵點匹配 ； 迴授控制

並列摘要

In recent years, unmanned aerial vehicle is getting more and more popular, for its convenient use and stable performance. People can take photos or videos with it at somewhere they cannot get. However, when there is only one person, and he wants to record the experience at that time, it will become more difficult. This thesis will try to solve this problem.
The first problem is to find the person, which is the pedestrian detection system in this thesis. The system uses deformable part model with HOG-SVM method, which can detect each part of the human body. By the reference of the standard statistics of human proportion, the corresponding parts can be matched if they belong to the same person. Hard examples which are generated from negative datasets are used to improve the accuracy of the detection. Several algorithms including detection rate and stability are introduced using deformable part model, and some optimization method can further improve the performance of the output. With several feedback mechanisms, the detection system can detect pedestrians with great accuracy.
After all the pedestrians on the input image are detected, a pedestrian tracking system is proposed to make the quad-rotor follow the target. An enhanced SURF key-point matching method is proposed to improve the matching quality, which can output more matching key-points and remove the obviously incorrect matching. The specific human can be tracked in static frame efficiently with a circulation region of interest, which is the output of the last frame to ensure the stability. Meanwhile, a feedback mechanism is introduced to make sure the system detects the pre-set target. Some lines on the screen are defined to give a reference of quad-rotor’s moving. If the target is out of these lines, the quad-rotor will move in corresponding direction, to ensure the target is in good position on the screen. While flying, the system will match with the 5 sides of the target’s key-points which are captured in advance, and compute the percentage of the front side among all sides. If the percentage is below a threshold, the quad-rotor will turn, and then fly to the front side of the target, to ensure the good tracking quality on the screen.

並列關鍵字

Quad-rotor ； Pedestrian Detection ； Object Tracking ； Trajectory Planning ； Key-points Matching ； Feedback Control

參考文獻

[1: Miller & Fels 2011]

[2: Khodaskar & Ladhake 2014]

A. A. Khodaskar and S. A. Ladhake, “Pattern recognition: Advanced development, techniques and application for image retrieval,” in Proceeding of 2014 International Conference on Communication and Network Technologies (ICCNT), Sivakasi, PP. 74-78, Dec. 18-19, 2014.

[3: Cox & Jager 1992]

G. S. Cox and G. de Jager, “A survey of point pattern matching techniques and a new approach to point pattern recognition,” in Proceeding of 1992 South African Symposium on Communications and Signal Processing, Cape Town, pp. 243-248, Sep. 11, 1992.