利用載具提供之空間關係以改善路面人形偵測之效能

Abstract

在智慧型車輛系統上，障礙物偵測一直是重要的議題，本篇研究認為除了一般行人之外，路面上騎著機踏車的人也需要偵測，因此將之納入考量。在過去的文獻當中，有許多與行人偵測的相關研究，這些研究大多採用梯度方向直方圖當成特徵，取得了不錯的成效，然而很少研究探討如何利用人與交通工具間的空間關係，且原先的行人偵測方面尚有一些改進空間。在改進一般的行人偵測方面，本篇論文提出了一個用物體本身紋理相似性的特徵，稱為Texture Self-similarity (TSS)，用來補足一般特徵較少描述的物體對稱及顯著特性。TSS是基於先前提出的LOP特徵來計算物體本身各區塊之間的紋理相似程度。在人形偵測實驗中，可以明顯看出TSS特徵改善了偵測的準確率。本篇研究也提出了一個對於機踏車偵測的強健特徵，利用輪胎的外型資訊來當成特徵，有效的進一步去判斷偵測到的機踏車是否正確，幫助提升整體系統之效能。 在做到有效的人形偵測及機踏車之後，本篇論文提出了使用空間關係來有效結合兩者之結果，在路面上偵測到的機踏車大多有駕駛者，同樣的偵測到駕駛者的附近應該也要有機踏車，故其空間關係應十分顯著。觀察之後使用駕駛者與機踏車的質心間位移間大小比例來當成所用之空間關係，分析此空間關係的分佈發現可以其呈現高斯分佈，故本論文利用高斯分佈的機率公式來計算其機率，假設變數間獨立故機率直接相乘，最終結果再過濾掉可能性低的組合。此外本篇論文也藉著空間關係來修正原本偵測到的視窗，混合理想位置與實際偵測位置以得到更精確的位置，藉由以上空間關係的結合運用進而提升偵測的結果。

Obstacle detection is an important area of intelligent transportation systems. We have to detect not only pedestrian but also on-road human with two-wheel vehicle. In the literature, there are a lot of work about pedestrian detection based on HOG feature and achieve good performance. However, there are few researches on how to use the spatial relationship and there is still room for pedestrian detection. To improve pedestrian detection, in this research we propose a new feature which is based on self-similarity, called Texture Self-Similarity (TSS). The main idea of TSS is to compute the similarity between blocks by comparing the histogram of texture information. The TSS also encode the symmetry and saliency of object, which is seldom be encoded. The experiments show that TSS feature is useful to improve the overall performance of system. We also design a strong feature for two-wheel vehicle detection. The feature is based on the shape of wheel to improve the accuracy of two-wheel vehicle detection. After improving the performance of human detection and vehicle detection, this thesis propose using the spatial relationship to combine these results. On-road human and vehicles are almost always appearing simultaneously. As a result, there is strong spatial relationship among on-road human and two-wheel vehicle. According to observations, we model the relative ratio and the displacement between the centroids as three variables. We use three independent Gaussian distribution to compute the probability by these variables. Then the candidates with low probability will be filtered out. Besides, we mix the ideal bounding box and detection bounding box to get better results. With the help of spatial relationship, we improve the system’s performance.