本論文主要是透過AI視覺辨識與立體視覺來辨識與定位麥皮蟲。使用麥皮蟲模擬一般蔬菜上的菜蟲,使用stereo-pi的左右攝像機取得的圖像為實驗樣本圖像。使用YOLOv4與Unet在stereo-pi實驗樣本圖像上辨識麥皮蟲並取得麥皮蟲在圖像上的位置。建立深度偵測模組讓YOLOv4預測框透過SAD與Unet產生的遮罩圖透過一維濾波能夠將左右圖片上的座標位置轉為視差而估算實際距離。訓練CNN深度預測模型來辨識Unet遮罩圖預測實際距離。評估指數分別為 YOLOv4 mAP@80 94.32%,Unet準確率99.2%,CNN深度預測準確率91.5%。
This research discusses how to detect and locate Zophobas by using AI visual identity and stereo vision.Use Zophobas to simulate vegetable worms on common vegetables.The sample images are obtained by the left and right cameras of stereo-pi.Identify and locate Zophobas from images by using YOLOv4 and Unet neural network model.Builddepth-detection module to estimate disparity from YOLOv4 bounding-box image with SAD and Unet mask image with one-dimensional filter.Use the disparity to calculate the distance between stereo-pi and Zophobas.Train a CNN module of Unet mask image to predict distance range.Evaluation accuracy:YOLOv4 mAP@80 94.32%,Unet validation accuracy 99.2%,CNN validation accuracy 91.5%.