EDGE AI ACCELERATION TECHNIQUES FOR REAL-TIME IMAGE PROCESSING IN EMBEDDED SYSTEMS

The article discusses the possible ways of accelerating Edge AI real-time image processing on embedded systems with limited computational and energy resources. The concentration is on machine vision, especially object recognition, and three platforms are examined: ESP32-Wrover-Dev v1.6, Sipeed Maix Bit, and NVIDIA Jetson TX2 DevKit. The ESP32 failed to gather enough power for full on-board AI processing, thus server-based recognition became necessary. The Sipeed Maix Bit, running the YOLOv2 model, was fairly successful, but the problems were in limiting the types of models and their accuracy over the distance. The NVIDIA Jetson TX2 with the models of SSD Inception V2, Peoplenet, and Peoplenet-Pruned tested by the Prestigio PWC420 HD camera and the Samsung Galaxy A54 camera was able to provide the highest recognition rates. Peoplenet gave the best results across the board, while Peoplenet-Pruned achieved similar accuracy with half the resource consumption. SSD Inception V2, however, appeared to behave erratically and perform poorly on longer distances or low-light conditions. The results suggest that for high-accuracy, real-time people detection on embedded systems, NVIDIA Jetson TX2 with the Peoplenet or Peoplenet-Pruned is the best choice, whereas lightweight platforms, such as the Sipeed Maix Bit, tend to be more consistent under varying light conditions yet less accurate for long-distance scenarios. The study highlights the compromises between performance and the consumption of resources versus conditions when realizing AI at the edge.