MDSSD-MobV2: An embedded deconvolutional multispectral pedestrian detection based on SSD-MobileNetV2

Multispectral pedestrian detection is known as a promising task in the field of computer vision and deep learning regarding its robustness in challenging conditions like adverse illumination, occlusion, and low-resolution imaging. Both the input data and the network layers can benefit from multispectral fusion. Due to the high computational cost of fusing visible and thermal modalities, it cannot be accomplished efficiently across many tasks in this domain, such as autonomous vehicles, robotic applications, security, and monitoring systems. Despite this fact, most recent efforts have focused on accuracy as a dominant parameter and have paid less attention to the importance of speed. As a solution, this paper proposes a fast Multispectral Deconvolutional MobileNetV2 based Single Shot Detector (MDSSD-MobV2) that effectively balances two crucial goals of accuracy and speed. This lightweight multispectral network is built upon the novel high-resolution MobileNetV2 followed by DSSD auxiliary layers, which fuse visible and thermal feature maps at multiple levels of the network architecture. This research is primarily dependent on developing a framework for a blind assistance detector that is light enough to run on embedded systems with low memory and processing power while still retaining the required accuracy. The evaluation results on the well-known KAIST benchmark show that this method attains sufficient speed on the Nvidia Jetson TX2 while still keeping up with recent solutions with a miss rate margin of less than 1%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\%$$\end{document} and improving their best speed by more than 3.5x on the non-embedded operating system of the GTX 1080Ti platform.