In this paper, we present a perception system for assisting robotic vehicles in smart cities, facilitating interaction with pedestrians, cyclists, and other motor vehicles while adhering to local traffic rules, all with the aim of enhancing traffic safety. Multiple Object Tracking (MOT) is a complex and fundamental problem in computer vision for robotic vehicles, requiring individual evaluation of various detected mobile agents to make informed decisions. To address this challenge, we utilize embedded and dedicated hardware systems, along with Deep Learning algorithms, as powerful tools for realtime processing of computer vision. In this work, we developed an Advanced Driver Assistance System (ADAS) with 91.85% (mAP) and 78.2% (IoU) accuracy for MOT using Nvidia’s Jetson-Nano and optimized the Deep-SORT YOLOv7 model in conjunction with the Kalman filter algorithm to achieve this capability, and a rate equal to or greater than 50% is already considered relevant for the task of detecting dynamic obstacles.
Real-Time Pedestrian Detection and Tracking System Using Deep Learning and Kalman filter: Applications on Embedded Systems in Advanced Driver Assistance Systems.
2023-10-09
6001068 byte
Conference paper
Electronic Resource
English
Real-Time Pedestrian Detection and Tracking at Nighttime for Driver-Assistance Systems
Online Contents | 2009
|Real-time pedestrian detection and tracking at nighttime for driver-assistance systems
Tema Archive | 2009
|Real-time pedestrian detection for driver assistance systems based on deep learning [10806-185]
British Library Conference Proceedings | 2018