Design, development, and performance evaluation of a robot for yield estimation of kiwifruit

Abstract

Abstract One of the applications of robotic farmer-assistant platforms equipped with machine vision systems is the evaluation of production yield before harvest without damaging the product. In this situation, farmers receive proper information for harvesting and post-harvesting management to decide about the required human resources, harvesting equipment, storage space, transportation, and product marketing. In this study, a machine vision system on a tracked vehicle was designed and developed for yield estimation of kiwifruit by traveling along the kiwifruit trellis. Several features, i.e., intensity histogram, the histogram of oriented gradients, shape context, and local binary pattern, were extracted from the images captured from the plants, and the number of kiwifruits in the images was predicted using the support vector machine (SVM). To improve the performance of the SVM, its parameters were optimized using evolutionary optimization methods, namely, particle swarm optimization (PSO), ant colony optimization (ACO), differential evolution (DE), and genetic algorithm (GA). The performance of the proposed method was compared with several deep learning techniques. The R2 of predicting the number of kiwifruits in the images was obtained equal to 0.96, 0.91, 0.73, 0.83, 0.90, and 0.63 for the proposed method, FCN-8S, ZFNet, AlexNet, GoogleNet, and ResNet, respectively. Furthermore, the results showed that the SVM enhanced with PSO exerted the highest area under the precision-recall curve compared to the deep learning methods. The findings of this study can be useful for the proper implementation of precision agriculture and the management of agricultural inputs.