基於藍牙與深度學習之溫室內定位及蔬果辨識

Abstract

近年臺灣農業受高齡化導致缺工問題日益嚴重，因此越來越多農業自動化的解決方案被用來解決人力不足的問題。過去於溫室中協助自動化採收的車輛多採用軌道的方式運行，雖能達到精準且穩定的移動，但會增添鋪設軌道的麻煩，因此本研究期望以藍牙無線通訊網路的方式對空間中的載具做定位，並使用深度學習的方式記錄栽培架上的果實數目和位置。本研究於可利用室內環境中作為參考座標的Beacon發射低功耗藍牙訊號讓載具做接收，載具則會依據接收到的低功耗藍牙的訊號強度指標(RSSI)的衰減程度做Beacon和載具之間的距離換算，通過接收到的多個訊號利用三邊定位的演算法計算載具於空間中的二維座標。溫室中果實則利用深度學習YOLO模型做物件偵測，而為了降低系統中硬體的運算量，本研究選擇較輕量化並可執行即時辨識能力的YOLO tiny模型，並使用247張牛番茄果實做YOLO tiny模型訓練。在室內定位精度上，使用訊號大小做定位精準深受環境中障礙物、硬體使用和參考座標密集程度的影響，在一維定位方面，短距離時之定位誤差可達公分級誤差，但隨著距離增加，定位誤差也隨之增加; 二維定位誤差除了受一維精準度影響，也會受Beacon密度的影響，在10個量測樣本點的試驗中，平均誤差為156.84 cm。而在應用深度學習的物件偵測成果上，使用189張測試影像對YOLO v4 tiny模型做結果評估，Precision為88.8%，Recall則為90.2%。相較於v3 tiny模型的成果，Precision小幅提升0.4%，而Recall則大幅提升了7%，且在高曝光程度的測試資料集之表現也有進步。

In recent years, the aging population has leaded to the labor shortage problem in Taiwan; as the result, more and more agricultural automation solutions are used to solve the insufficient manpower problem. In the past, the vehicles that assisted automated harvesting in the greenhouse are mostly operated by rails. Although highly accurate position and stable movement could be achieved, it would increase the cost and time of laying the track. Therefore, this study aimed to find the vehicle position in the greenhouse by means of Bluetooth wireless communication network, and use deep learning algorithm to record the number and position of fruits on the cultivation racks. In this study, Beacons as reference coordinates in the indoor environment were used to transmit Bluetooth Low Energy (BLE) signals for the vehicle to receive. Then, the vehicle would convert the distance between the Beacons and the vehicle according to the Bluetooth signal ‘s Received Signal Strength Indication (RSSI) and use the trilateration algorithm to calculate the vehicle ‘s two-dimensional coordinate. YOLO was used as the deep learning algorithm for fruits object detection in this study. In order to reduce the computational load of the hardware in the system, YOLO tiny with lighter model that can perform real-time identification was selected and used 247 tomatoes images filming in the greenhouse for YOLO model training. In terms of indoor positioning accuracy, using RSSI for positioning accuracy is greatly affected by obstacles in the environment, the use of hardware and the density of reference coordinates. For one-dimensional positioning, at short distances, the positioning error could reach centimeter-level errors. As the distance increased, the positioning error also increased. For two-dimensional positioning, the estimated error is affected by one-dimensional accuracy and the density of Beacons as well. In the test of 10 measurement sample points, the average estimated error is 156.84 cm. In terms of target detection results using deep learning, using 189 test images to evaluate the results of the YOLO v4 tiny model, the Precision rate was 88.8% and the Recall rate was 90.2%. Compared with the results of the YOLO v3 tiny, Precision had increased slightly by 0.4%, while Recall had increased by 7%, and the performance in the high-exposure test data set has also improved.