基於深度學習與視覺之車輛偵測

Deep Learning and Vision-Based Vehicle Detection

黃登淵(DENG-YUAN HUANG)

《科學與工程技術期刊》 15卷2期 (2019/09) Pp. 1-11

摘要

本文提出植基於Yolov3架構之深度旋積神經網路以達成端到端之車輛偵測與追蹤。由於利用數量龐大的ImageNet影像資料庫重新訓練整個深度網路能提升的準確度相當有限，因此採用微調方法是較為可行的做法。配合微調方法，本文自製車輛資料庫，其中包含巴士、汽車及卡車等三種不同型態的車輛。在微調階段預訓練過程中，本文所提方法可以達成98%的分類水平。在車輛偵測方面，本文採用4種不同情境的影片，包含有高速公路、巷弄街道、夜間車道與市區道路等，達成的偵測率分別為78.8%、69.1%、86.2%及88.1%。另外，在輸入影像解析度為1920×1080像素以及基於CPU執行的情況下，最終偵測速率每幀可達420-470ms之水準。最後，總體的平均偵測率與誤判率分別為82.1%與9.8%，足以驗證本文所提方法之可行性與有效性。

關鍵字

深度學習 ； 旋積神經網路 ； 車輛分類 ； 車輛偵測

並列摘要

This paper proposes a deep convolutional neural network (DCNN) based on the YOLOv3 architecture to design an end-to-end vehicle detection and tracking system. It is considered feasible to use the fine-tuning method because the increase in detection accuracy of vehicles is quite limited even through retraining the entire DCNN using numerous images in the ImageNet dataset. To facilitate use of the fine-tuning method, we designed a custom database that includes three different types of vehicles, namely buses, cars, and trucks. In the pre-training phase of the fine-tuning method, the proposed method achieved a classification rate of 98%. In vehicle detection, we used four test videos with different scenarios, namely highways, alleyways, night lanes, and urban roads, to achieve detection rates of 78.8%, 69.1%, 86.2%, and 88.1%, respectively. The CPU only-based detection speed can reach 420-470 ms per frame when the input image size is 1920 × 1080 pixels. The overall average detection rate and false alarm rate for the four test videos was 82.1% and 9.8%, respectively, which indicates the feasibility and effectiveness of the proposed method.

並列關鍵字

deep learning ； deep convolutional neural network (DCNN) ； vehicle classification ； Vehicle detection