پياده‌سازي بهينه الگوريتم تشخيص اشياء بر روي سخت‌افزار مبتني بر ZYNQ

Object detection has been one of the fundamental tasks in computer vision, focusing on the identification, localization, an‎d classification of objects in images an‎d videos. Its applications include autonomous vehicles, medical imaging, an‎d video surveillance, among others. The deep learning-based approach to object detection includes single-stage methods like YOLO an‎d SSD for real-time applications an‎d two-stage methods like Faster R-CNN with higher accuracy. The deep learning object detection networksʹ hardware platform implementation, based on ZYNQ chips with the advantages of low energy consumption an‎d parallel processing capabilities designed using FPGA technology, suffers from great challenges. Some of these include: limitations in storage an‎d computing resources, energy consumption management, an‎d model accuracy preservation. This work discusses the deployment of the single-stage object detection model YOLOv7-Tiny on ZYNQ chip-based hardware with the VisDrone dataset. Towards the encountered problems, a model compression technique quantization has been performed. However, quantization resulted in more false positive bounding boxes an‎d thus decreased model precision by up to 5 percent an‎d reduced the visual quality of the results produced by the model. In this regard, local information of the predicted bounding boxes an‎d the Canny edge detection filter are used to refine the predicted bounding boxes. The results of this research demonstrate that this correction method has improved the modelʹs precision by 2 percent.

5