A deep learning approach for object counting on embedded systems

This thesis discusses the implementation of a piece counter, based on deep learning methods, to be used on industrial packaging lines. In particular, we aim at obtaining a system capable of detecting small objects (diameter < 1cm) and that can be executed on embedded devices. The object detection models SSD MOBILENET V2 and SSD RESNET 50 , after a brief analysis of their characteristics, result to be the most suitable for this application and their performance is measured by using bolts of different sizes as test object to be detected. In the experimental results, the pros and cons of the two models are analyzed in terms of accuracy, inference time and efficiency. All tests are performed on the development board NVIDIA JETSON NANO in order to optimize models using TensorRT and evaluate the results.