Real Time Performance Comparison of Multi-class Deep Learning Methods at the Edge

Abstract

Nowadays, machine learning algorithms are used intensively for the detection and identification of various entities on an image with high level of accuracy. Deep learning provides even higher performance results in image processing. However, deep learning requires high computing effort for classification and thus limiting the usability of method in AI at the edge especially in the presence of different classes. In this paper, deep learning is applied as a solution to the problem of detecting and identifying more than one entity of different classes in the image in real time at the edge. In the study, a multi-class traffic sign data set was used. The data set consists of six different sign classes and 6000 high resolution images (1280 × 720). Performance of two models, DetectNet and MobileNet (MobiNet) models were compared. MobileNet had better accuracy compared to DetectNet since it has higher number of parameters. But high number of parameters lessens the real time performance of the model at the edge. According to the tests performed on the Nvidia Jetson TX2 embedded system, the DetectNet model can provide 6–7 FPS performance while the MobileNet 160 version of MobileNet model with $\propto \; = \;0.25$ has achieved only of 1–2 FPS.