A Comparative Analysis of Deep Learning-Based Object Detectors for Embedded Systems

The paper presents a comparative analysis of popular deep learning-based object detectors, focusing on evaluating their real-time performance on embedded platforms. The objective was to identify the optimal algorithm that can efficiently operate within limited computational resources and memory constraints of embedded systems while minimizing energy consumption. We tested a set of popular detectors, such as Faster R-CNN, YOLO, and SSD MobileNet and evaluated their performance on the testing platform with CUDA cores. A faster variant of bounding box suppression, blazing fast Non-Maximum Suppression (NMS), was employed to enhance efficiency. Additionally, the generalized intersection over union (GIoU) metric was adopted as the optimal testing method, allowing for more accurate object localization and detection. Results demonstrate that the SSD MobileNet-V2 and Tiny YOLO detectors exhibit promising real-time performance on embedded platforms, although their accuracy is diminished compared to larger models.