Abstract:
Pesticide spraying is one of the most important farm activities related to the protection of plants. The application of pesticides by unmanned aerial vehicles (UAV) can effectively avoid the harm caused by pesticides to human body, but the amount of pesticides is limited and it is difficult to locate the pesticides between flights. The unmanned navigation tractor with GNSS (global navigation satellite system) has realized the automatic navigation of planting and cotton harvesting in large area cotton field while could not work in cotton plant protection operation because the complex crop growth environment on site. Visual navigation vehicle can carry a large amount of pesticide, and its operation path is determined by the growth state of crops in the farmland avoiding crushing seedling. Therefore, it has great potential for the unmanned plant protection vehicle based on visual navigation in the field environment. Nowadays, sprayers mounted on tractors have being utilized in cotton protection in Xinjiang. The present studies for objects mainly focus on static images, and the dynamic detection of crop rows in video sequence is studied barely. Aiming at that, A plant protection unmanned vehicle was exploited, which can enhance and complement the intelligent agriculture. After the vehicle finish spraying in one cotton row, it will stop safely and accurately just at the edge of the proposed position. Subsequently, its 4 wheels will rotate 90° simultaneously, so as to ensure it can move in a vertical direction through the cotton rows; a camera on the opposite side can also be utilized to count the cotton rows and avoid repeated spraying. When the vehicle passes cotton rows preseted, its 4 wheels can simultaneously rotate 90°again along the same direction to prepare for the next spraying. Then a video-based dynamic count method for cotton rows is proposed to determine the interval distance of linear operation area in this paper. By tracking the centroids of cotton plant, the number of cotton rows could be counted in real time. Firstly, the color image from video became grayscale to emphasize information of cotton rows with prominent green component by calculation of 2G-R-B. Then, a region of interest (ROI) was set not only to reduce the calculation amount, but also to avoid the lacking of seedling in the end of cotton rows. Secondly, the vertical cumulative histogram of gray in ROI was solved and the histogram vs ROI width curve was obtained. The wave crests of the curve determined by two conditions were found, in which the width and the sharpness should be satisfied with threshold. Then a floating window was established to positioning the cotton row. Besides that, the grayscale centroids of each floating window were calculated, which represent the cotton rows where each one was located. Thirdly, a target window with the size of 65×65 was established for each centroid. With the growth of cotton, the branches and leaves intertwined in which situation, the gap between the 2 cotton rows became smaller, and the cotton leaves of one row were very loose. In this case, one cotton row might correspond to more than one target window. To avoiding the influences of the growth, the overlap rate was calculated between each 2 adjacent target windows. And the 2 target windows whose overlap rate meeting the condition (large than 0.1) should be merged together. It had been done until no overlapped rate meeting the conditions. Thus, each cotton row had been identified. For tracking, starting from the second frame, the overlap rate between each target window and other target windows from previous frame was judged, if it met the condition, the two target windows in the two successive frames were connected. Then the cotton row could be tracked. The test proved that the processing time of detecting cotton row was approximately 150 ms per frame, which satisfied the requirement of practical application for the cotton protection in the field.