引入環境資訊之路上物體偵測與追蹤

Abstract

因為有各式各樣的挑戰，使得能夠穩定偵測在影片中的行人是困難的。在偵測行人的各種鑑別式特徵中，其中一個最成功的當屬方向梯度直方圖(Histogram of Oriented Gradients, HOGs)。雖然主要的輪廓資訊成功的在HOG中描述，背景雜亂分佈擾亂了梯度資訊。因此，一個基於HOG的擴充，名為細粒方向梯度直方圖(Histogram of Oriented Gradient of Granules, HOGGs)進而被提出。相較於HOG對於每個像素計算梯度，HOGGs 計算小區域之間的梯度。背景雜亂問題因此可因為額外的區域資訊而解決。 此外，一個穩健的多車輛與多車道偵測並且同時整合車道與汽車資訊之系統在論文中開發。大部份的研究目前只各別分開偵測車道或是汽車。為了達成更加可靠的結果，車道線與汽車之間可相互支持的關係應該被建模考量。基於此，本論文透過機率資料關聯濾波器(Probabilistic Data Association Filter)整合車道與汽車之空間與時間資訊。因為結合汽車與車道能夠改善汽車與車道追蹤的一致性，因而改善汽車偵測的效能。在實驗中也驗證了所提出的系統可以可靠有效的偵測多車道與多車輛。 在影片中穩定的同時偵測路上行人與車輛亦是具有挑戰性的問題。一些文獻主要只偵測或追蹤單一種目標物。然而，系統偵測或是追蹤效能可以透過幾合資訊的幫助而有效改善。因此，在本論文中不是分別偵測行人或是汽車，而是提出了結合先前所述之整合異質性資訊的框架。在此，行人與汽車偵測很自然的透過了場景幾合資訊結合。相機的俯仰角透過的創新的消失線估測法而取得。不採用傳統透過長直線交點做為消失點之方法，被追蹤的道路上物體的資訊被拿來參考。具體而言，每個被追蹤的路上汽車或是行人會對可能的消失線位置投票。因此，消失線的位置即使在雜亂或是擁擠的背景中仍可被判斷。因此場景幾合資訊可以在有挑戰性的環境中被估計。透過貝氏網路，此資訊可以幫助改善偵測結果。最後，為了驗證效果，在KITTI資料集中作了大量的實驗。目前在該資料集中表現領先的偵測器Regionlet偵測器，可以因為本方法得到不錯的改善。

To detect people in a video sequence robustly is hard due to various challenges. One of the most successful discriminative features for finding people goes to the Histogram of Oriented Gradients (HOGs). Although the major contour information is encoded in the HOG feature well, background clutter disturbs the gradient information. Thus, an extension of the HOGs, called histogram of oriented of gradient of granules, is proposed. Instead of collecting gradient information over each pixel, the histograms of gradients of small regions are computed. The clutter background problem can be solved by encoding extra region information. A robust system for detecting on-road multiple vehicles and multiple lanes while integrating both lane and vehicle information is designed. Most researches so far can only detect single/multiple lanes or vehicles separately. To achieve more reliable results, the relationship between lane and vehicle which can support detection of either of them should be modeled. Following this, we thus integrate spatial and temporal information of lanes and vehicles through employment of the probabilistic data association filter model. Such integration will improve the consistency of vehicle and lane tracking, and hence increase the performance of on-road vehicle detection. The experiments have validated our hereby proposed system for detecting multiple vehicles and multiple lanes satisfactorily and reliably. To robustly detect people and vehicle on the road in a video sequence is also a challenging problem. Most researches focus on detecting or tracking of specific targets only. Nevertheless, the performance of the system conceivably can be improved with the help of the geometry information. Thus, in this research, instead of detecting vehicle or pedestrian individually, a framework integrating the aforementioned heterogeneous information is proposed. Here, our approach let the system naturally integrate different information using the scene geometric information. The camera’s pitch angle is estimated with a novel vanishing point estimator. Instead of detecting the vanishing points using line intersection approach, the object information from tracker are also considered. Specifically, the detected vehicle or pedestrian will cast votes for the hypothesized horizon line. The vanishing line can be detected even when the scenes are cluttered or crowded, and thus the geometric information can be estimated under challenging circumstance. Such information of scene can help the system refine our detection results through Bayes’ network. Finally, to verify the performance of the system, comprehensive experiments have been conducted with the KITTI dataset. It is quite promising that the state-of-the-art detector, in our case, Regionlet detector, can be improved.