影像處理應用於葉菜收穫機行走與採收之研究

根據農委會的統計，在臺灣五十歲到六十四歲的農業勞動力人口佔了全農業勞動力人口的44.5%，此數據突顯了農業人力結構的問題，再加上臺灣農業人口的高齡化以及農村人口外流使得作物採收期雇工不易，使得作物無法預期收成。若錯過採收期將使得作物的品項不佳，嚴重者甚至枯黃。在目前還沒有開放農業外勞人口的政策下，臺灣需透過農業自動化來改善農業缺工問題。本研究主要目的為研製出一台能夠在溫網室內利用影像處理進行行走與採收之葉菜採收機。實驗分為三個部分，分別為尋找葉菜適當的影像處理法，履帶車割刀的行為控制以及履帶馬達的行走控制。割刀馬達控制的部分先針對40張在臺灣大學拍攝的作物照片進行影像處理，每張照片再分成作物區與土壤區。經由作物區的白點數計算，來估計作物應有的白點數量。最後執行大樣本、母體變異數未知的臨界值檢定，找出代表作物的臨界值，來決定割刀是否應該作動。實驗結果顯示，當相機畫素為4068*3456時，作物區面積大於5825540，有95%信心水準為待收割作物。履帶馬達的行走控制部分，使用一輛自製履帶車，連接Raspberry Pi並裝上網路攝影鏡頭與馬達驅動板，來控制整台履帶車的行走。實驗中將網路攝影機拍攝之圖像進行HSV色彩空間轉換，再進行二值化後，利用程式計算網路攝影機左右兩側作物區之面積差異，透過PID參數的調整來控制馬達轉速，達成最小的穩態誤差與理想的最大超越量。實驗結果顯示，當kp=0.2時，可以達成最好的循跡效果。

關鍵字

影像處理；機器視覺； Raspberry Pi ；農業

並列摘要

According to Council of Agriculture, R.O.C. (Taiwan), those aged between 50 and 64 years account for 44.5% of Taiwan’s total agricultural labor force. This figure highlights Taiwan’s problematic agricultural labor force structure. Because of the aging agricultural labor force and increasing urbanization, finding workers to hire during the harvest season has become challenging. Accordingly, crops cannot be harvested in time, so crop quality has deteriorated. Additionally, crops have become increasingly likely to fail. Because current policy does not allow foreign agricultural workers to work in Taiwan, agricultural automation is required to resolve this problem. In this study, we developed a leafy vegetable harvester that can move within a screened greenhouse and harvest leafy vegetables through image processing. The experiment comprised three parts, including the best image processing method I should use, development of behavioral control for the crawler-track cutter and development of movement control for the crawler-track motor. To develop appropriate motor control for the cutter, 40 images of crops photographed in National Taiwan University underwent image processing. Each image was divided into crop sections and soil sections. The number of white dots in the crop sections was calculated to estimate the number of white dots that crops should have. Finally, the critical value method was adopted for a large sample size and unknown population variance to ascertain a representative crop’s critical value and to determine whether the cutter should take actions. The results revealed that when the camera resolution was 4068 × 3456 and the area of the crop section exceeded 5,825,540 pixels, with a confidence level of 95%, the area reflected crops to be harvested. For movement control of the crawler-track motor, a crawler-track car produced by this study was utilized. A Raspberry Pi, a web camera, and a motor driver board were installed to the car to control and monitor its movement. In the experiment, images obtained through the web camera were processed through hue-saturation-value color-space conversion. Subsequently, image banalization was conducted, and software was used to calculate the differences in area between the two crop sections at the right-hand and left-hands side of the image. The motor’s rotational speed was controlled by adjusting the proportional-integral-derivative parameters to achieve minimal steady-state errors and optimal maximal overshoot. The results revealed that the optimal tracking effect was produced when kp = 0.2.