基於FPGA之無人水面載具影像分割系統

Abstract

於無人水面載具技術發展中，自動導航扮演了重要的角色。為了利用電腦視覺技術實現無人水面載具的自動導航，視覺系統必須具有分辨障礙物的能力。影像分割是種將像素點進行分類的影像處理技巧，於無人水面載具應用中，可透過影像分割將輸入影像分割為水體、天空、障礙物，藉以供導航使用。在無人水面載具的應用場景，視覺系統必須符合即時性、精準性、低功耗、低成本的需求。本研究利用FPGA作為硬體平台，開發適用於無人水面載具的即時影像分割系統。利用包含1325張真實水域影像的MasTr1325資料集，訓練了用於水域影像分割的卷積神經網路，並利用Vitis-AI架構佈署至FPGA，實現神經網路的硬體加速。而為了改善系統效能，我們透過可程式化邏輯對影像前處理運算進行硬體加速，並使用多執行緒設計應用程式。實驗結果顯示，本研究提出之系統吞吐量可達25 FPS，而在包含8175張影像之MODS測試資料集上，達到超過85%的F1分數，可以滿足無人水面載具導航之需求。

In the development of unmanned surface vehicle (USV) technology, autonomous navigation plays a critical role. To achieve autonomous navigation using computer vision technology, the visual system of the vehicle must be capable of distinguishing obstacles. Image segmentation is an image processing technique that classifies pixels into meaningful regions. In the application of autonomous ship, the input image can be segmented input into water, sky, and obstacle for navigation. During USV navigation, the visual system must meet the requirements of real-time performance, precision, low power consumption, and low cost. This thesis used field programmable gate array (FPGA) as the hardware platform to develop a real-time image segmentation system suitable for USV application. We trained a convolutional neural network (CNN) with the MasTr1325 dataset for maritime image segmentation and deployed it to FPGA using the Vitis-AI framework to achieve hardware acceleration. To further improve system performance, we accelerated the image preprocessing step through programmable logic and used multithreading in the application design. Experimental results showed that the proposed system achieved a throughput of 25 FPS, and the segmentation results reached an F1 score over 85% on MODS dataset containing 8175 images, meeting the requirements of USV navigation.