Real-time Object Detection and Semantic Mapping on CPU-Powered Mobile Robot

Deploying deep learning object detection models on resource-constrained mobile robots is a significant challenge because these models typically require power-hungry GPUs for high-performance inference. This requirement creates a critical bottleneck for mobile robots equipped only with CPU-based computers, such as Raspberry Pi 4B. We propose a semantic mapping system integrating an int8-quantized YOLOv10p model with ROS2 and SLAM. The system runs the detection node projecting 2D detections from a monocular camera into a map using a ground-plane assumption. Our real-world experiments provide the critical performance baseline. The system achieves 100% Recall. However, this system suffers from low 54.5% Precision. We demonstrate through qualitative analysis that the low precision is caused by clusters of False Positives, stemming from the instability of the monocular 3D projection method and inaccuracy in estimation of object detection. Our work confirms the feasibility of integrating YOLOV 10 into the CPU-only ROS2 system.