基于深度学习与目标跟踪的苹果检测与视频计数方法

基于机器视觉技术自动检测苹果树上的果实并进行计数是实现果园产量测量和智慧果园生产管理的关键。该研究基于现代种植模式下的富士苹果视频，提出基于轻量级目标检测网络YOLOv4-tiny和卡尔曼滤波跟踪算法的苹果检测与视频计数方法。使用YOLOv4-tiny检测视频中的苹果，对检测到的果实采用卡尔曼滤波算法进行预测跟踪，基于欧氏距离和重叠度匹配改进匈牙利算法对跟踪目标进行最优匹配。分别对算法的检测性能、跟踪性能和计数效果进行试验，结果表明：YOLOv4-tiny模型的平均检测精度达到94.47%，在果园视频中的检测准确度达到96.15%；基于改进的计数算法分别达到69.14%和79.60%的多目标跟踪准确度和精度，较改进前算法分别提高了26.86和20.78个百分点；改进后算法的平均计数精度达到81.94%。该研究方法可有效帮助果农掌握园中苹果数量，为现代化苹果园的测产研究提供技术参考，为果园的智慧管理提供科学决策依据。

Apple detection and counting using real-time video based on deep learning and object tracking

Abstract: Anaerobic digestion has been widely utilized to dispose of agricultural organic wastes. The renewable energy of methane can be produced during the treatment, together with the digestates rich in the nutrients for the fertilizer. However, the ammonia nitrogen can be tended to accumulate during anaerobic digestion, when using a large proportion of protein-rich substrates, such as chicken manure, pig manure, and kitchen wastes. Once the concentration of ammonia nitrogen reaches over 3 000 mg/L in the anaerobic process, the ammonia inhibition is likely to happen, resulting in the decrease of microorganisms'' activities and methane production during anaerobic digestion. Struvite precipitation can be a useful way to remove the ammonia nitrogen and phosphorus in the digestates and wastewater. Many studies have been reported to optimize the reaction conditions, such as the molar ratio of Mg to P, pH level, and temperature, to recover the struvite. However, there are only a few studies to combine struvite precipitation with anaerobic digestion. This study aims to investigate the effect of in-situ struvite precipitation on the anaerobic digestion of chicken manure. The MgCl2·6H2O and K2HPO4·3H2O were mixed into the feeding substrate in the stable running reactors for 6-7 consecutive days to remove NH4+-N. The theoretical removal rate was at the speed of 3 000 mg/d. Some parameters were detected, including the concentration of ammonia nitrogen, methane yield, total volatile fatty acids (TVFA), and pH during anaerobic digestion. After the first operation of adding MgCl2·6H2O and K2HPO4·3H2O, the concentration of ammonia nitrogen and TVFA were reduced from 2 937 to 1 466 mg/L, and 2 317 to 72 mg/L, respectively, whereas, the methane production was 0.39 L/gVS increased by 18%, compared with the control group (0.33 L/gVS), where the utilization rate of magnesium and phosphate was 91%. After the second operation, the concentration of ammonia nitrogen and TVFA were reduced from 2 232 to 762 mg/L, and 2 321 to 25 mg/L, respectively, whereas, the methane production was 0.33 L/gVS increased by 10% approximately, compared with the control group (0.30 L/gVS), where the utilization rate of magnesium and phosphorus was 90%. The results demonstrated that the addition of exogenous MgCl2·6H2O and K2HPO4·3H2O greatly contributed to mitigating the ammonia inhibition by struvite precipitation during the anaerobic digestion. An optimum pH was 8.5-9 for the struvite precipitation in the nutrient recovery of wastewater. A high utilization rate of magnesium and phosphorus was also achieved, when the pH of the system was 6.9-7.8, due to the high ammonia nitrogen concentration in the system. As such, it can be widely expected to promote the struvite precipitation to consume most of the magnesium phosphate salts. The exogenous MgCl2·6H2O and K2HPO4·3H2O can release H+ in the system, when the struvite was formed the lower pH to consume the alkalinity in the digester, easily leading to the acidification of anaerobic digestion. Consequently, the amount of exogenous MgCl2·6H2O and K2HPO4·3H2O needs to be controlled within a reasonable range for the stable anaerobic process.