Iot-Yolox: Comparative Evaluation And Advancement Of Object Detection Models For Iot-Based Robotic Arms

Abstract

The object and image detection capabilities emerging from IoT systems with robotic arms promote automation across healthcare institutions and agricultural settings and supply chain management as well as production facilities. Real-time decisions and robotic system intelligence depend on object detection since it serves as their primary essential foundation. A robotic system uses this detection process as part of its methodology to identify objects which allows robots to connect with objects around them. The implementation of advanced deep learning object detectors in IoT environments becomes challenging because of constrained memory resources combined with limited computing capability and restricted bandwidth and necessary time-sensitive operation requirements.The study primarily investigates the relationship that exists between the performance capabilities of leading detection algorithms and IoT robotics computational boundaries. We have analyzed seven current object detection algorithms with YOLOv5, SSD, Faster R-CNN, MobileNet-SSD, EfficientDet, RetinaNet together with CenterNet. The evaluation framework analyzes seven object detection models through precision metric tests of mean Average Precision and speed metrics of frames per second as well as measuring model complexity by floating-point operations per second and energy efficiency parameters along with system memory consumption and response time during real-time operation. The included mathematical explanation accompanies architectural design specifications to explain function and application compatibility for each model.The project presents IoT-YOLOX as a newly developed lightweight detection model specifically made to solve IoT-based robotic challenges. The IoT-YOLOX model uses YOLOX architecture and EfficientNetV2 for efficient feature extraction and PANet++ module for advanced multi-scale feature aggregation and the Quantization Aware Training method achieves accuracy retention during quantization. TensorRT understands the NVIDIA Jetson Nano and Google Coral TPU for performing edge AI acceleration in real time.Experimental results show IoT-YOLOX produces better performance results than baseline models because it achieves high accuracy ratings with fast processing along with streamlined resource usage and power requirements. Embedded hardware support at 62 FPS enables the IoT-YOLOX model to process Pascal VOC dataset mAP of 81.2% while operating faster than YOLOv5 by more than 30%. This model demonstrates perfect fit for IoT real-time applications through its capabilities to maintain both high-speed operations and power-efficient functioning and real-time response ability.The research provides three primary contributions which include a profound evaluation of IoT robotic arm detection models, the creation of the edge-optimized IoT-YOLOX model and experimental performance assessment for all evaluation criteria. The research combines a detection model selection guide for IoT conditions with an operational robotic solution.