An Energy-efficient and Accurate Object Detection Design for Mobile Applications

Deep Convolutional Neural Networks (DCNNs) are imperative to state-of-the-art computer vision algorithms. In spite of the attractive qualities of DCNN s, they have been excessively expensive to be applied on large scale high-resolution images and video sequences. In order to implement DCNN models on edge platforms, we tend to optimize the DCNN model by considering energy efficiency and detection accuracy simultaneously. In this paper, we analyze the energy consumption, detection accuracy, and execution speed of our model and those of the state-of-the-art models based on a mobile platform called Jetson Nano. We adopt the performance index from Low Power Computer Vision (LPCV) challenge which considers power, mAP and FPS at the same time to evaluate these models in an overall point of view. On Jetson Nano, the presented system boosted with the GoP-mode technique can achieve an execution speed of near 20 frames per second, and high mean average precision of 59.9% under MS COCO test sets. Compared with the state-of-the-art models, e.g., YOLOv5, the LPCV score improves as high as 76.33%. If the GoP-mode acceleration is included, the LPCV score of Agilev4 reaches even 90.6 times of that ofYOLOv5.