Classification of targets detected by mmWave radar using YOLOv5

Mohamed Lamane, Mohamed Tabaa, Abdessamad Klilou

Procedia Computer Science · 2022 · 12 citations · 8 references

DOIFull text

Open access

Concepts

Abstract

Frequency Modulated Continuous Wave (FMCW) radar is gradually used in industrial applications or advanced driver assistance systems (ADAS) for obstacle detection, this is due to its low cost and extreme operating conditions, unlike camera and Lidar, which is more accurate than ultrasonic or infrared sensors. In this work, we used the dataset of a cascaded radar called ColoRadar, we transformed the dataset from the polar representation (range, angle) to a Cartesian representation (X, Y), then we normalized the dataset so that all the Doppler peaks are between 0 and 1, to obtain uniform heatmap images, in which to train a YOLOv5 model that can classify and detect the presence of a human on heatmap images, we propose a YOLOv5 model and then compare it with several YOLOv5 models, the average accuracy (mAP_O.5) reaches 0.89, and the computational power consumed by this model is 12 Billion Floating Point Operations Per Second (BFLOPS).

References

8