Implementation of deep learning models for object recognition on embedded computing platforms

U okviru rada odabrane dvije ugradbene računalne platforme: Google Coral Dev Board i NVIDIA Jetson Nano. Dan je uvid u značajne karakteristike odabranih platformi te su obje uspostavljene za rad. Na njih se zatim implementiralo nekoliko modela dubokog učenja za detekciju objekata. Sami modeli, bazirani na različitim varijantama EfficientDet-Lite modela, istrenirani su na vlastitom skupu podataka, nakon čega si i evaluirani. Treniranje modela provedeno je primjenom tehnike prijenosa učenja. Istrenirani modeli zatim su prilagođeni za izvođenje na pojedinoj platformi. U slučaju Coral Dev Board-a, prilagodba je podrazumijevala potpunu 8-bitnu kvantizaciju modela te njihovu konverzija u TensorFlow Lite format. S druge strane, prilagodba modela za Jetson Nano odnosila se na njihovu konverziju u TensorRT format. Nakon uspješne prilagodbe, modeli su implementirani na ugradbene platforme i njihov rad testiran je korištenjem USB kamere. Uz to su testirane i performanse pojedine platforme, mjerenjem prosječnog vremena inferencije svakog modela, te su dobiveni rezultati uspoređeni.In the context of the thesis, two embedded computer platforms were selected: the Google Coral Dev Board and the NVIDIA Jetson Nano. An insight into the significant characteristics of the selected platforms is given, and both were set up for work. Several deep learning models for object detection were then implemented onto them. The models themselves, based on different variants of the EfficientDet-Lite model, were trained on a custom made data set, after which they were evaluated. Model training was carried out using the transfer learning technique. The trained models were then modified to run on the selected platforms. In the case of the Coral Dev Board, the modifications involved full 8-bit integer quantization of the models and their conversion to the TensorFlow Lite format. On the other hand, the modifications of the models for the Jetson Nano refered to their conversion to the TensorRT format. After this was done, the models were implemented onto the embedded platforms and were tested using a USB camera. In addition, the performance of each platform was tested by measuring the average inference time of each model, after which the obtained results were compared