DETECTION OF FRUITS USING EDGE AI APPLICATION BASED EMBEDDED SYSTEMS

On several edge AI devices, an in-house fruit detection system was developed and tested. In order to increase its precision in the presence of small and mainly occluded materials, the classical Yolo architecture has been updated and adapted. It was trained with a personalized data set and checked with all available OIDv4 photos of a real orchard. The accuracy reports have seen an increase in recall and accuracy for targets of different scales. Experimental assessments have been carried out to demonstrate the efficiency achieved by the various embedded systems chosen in terms of inferior speed and power consumption. Expectations of using the tested method have been encouraging for the production of real-time positions and detection numbers with limited electricity consumption. For different uses, from fruit counting, health assessments and intelligent packing, a full system may be integrated into the study. In fact, further work only focuses on the estimated yield using the methods suggested to accurately count the number of fruits.