Edge-Based Deep Learning for Traffic Object Detection in Driving Scenarios

Abstrak

Object detection is an important task in autonomous driving. Accuracy and processing speed often become a trade-off in edge deep learning-based object detection implementations. This study aimed to comparatively analyze the performance of deep learning-based object detection methods at the edge in detecting traffic-related objects in driving scenarios. This study contributes to the field by implementing fine-tuning methods on a custom dataset, on driving scenarios in Yogyakarta, Indonesia, and deploying them onto a resource-constrained edge computing device. PyTorch is used as the deep learning framework, with TorchVision as the main library, and Open Neural Network Exchange (ONNX) is used to convert PyTorch models into a standardized graph for edge implementation. The experimental results show that the two-stage detector Faster R-CNN ResNet50 with an input size of 800 outperforms other methods, achieving the best mAP of 37.6% and mAPS of 21.1%, but its inference time was the longest, reaching 78.89 s. In contrast, Faster R-CNN MobileNetV3-Large with an input size of 320 achieved the fastest inference time (0.58 s) with significantly lower mAP, mAP0.5, and mAP0.75 (11.2%, 19.5%, 11.9%, respectively). Conditional DETR achieved a moderate inference time (17.31 s) and considerable mAP (32.7%). FCOS with an input size of 800 had a higher mAP than Conditional DETR (34.8% vs. 32.7%) and an inference time twice as fast as Faster R-CNN ResNet50 (38.92 vs. 78.89 s). Therefore, Conditional DETR and FCOS are preferable for resource-constrained edge computing implementations.

Identifier

• OpenAlex: https://openalex.org/W7163774356
• DOI: https://doi.org/10.48084/etasr.17760