3D Bounding Box Estimation Using Deep Learning and Geometry Based on Yolov7 Output on Single-Board Computer

This study investigates the enhancement of 3D bounding box estimation techniques for object localization on SingleBoard Computers (SBCs), focusing on the Jetson Nano platform. The adaptation and optimization of deep learning models, specifically transitioning from VGG networks and YOLOv3 to more efficient alternatives like MobileNetV3 and YOLOv7, within the constraints of SBCs. The implementation leverages the advanced capabilities of MobileNetV3 for 3D bounding box generation, coupled with the superior detection accuracy and speed of YOLOv7 for object detection. The research employs an innovative loss function to improve 3D orientation predictions and utilizes geometric constraints from 2D bounding boxes for precise object localization. A comparative analysis of MobileNet V3, VGG-19, and MobileNet V2 models on the Jetson Nano inference speed and consistency reveals that MobileNet V3, optimized using TensorRT, significantly outperforms others, a preferable candidate for solutions in real-time environments. The study concludes that the strategic optimization of deep learning models on SBCs, like the Jetson Nano, markedly enhances the performance and applicability of 3D bounding box estimation in edge computing environments, offering valuable insights for deploying advanced object detection technologies in resource-constrained scenarios.