Autonomous vehicles multi obstacle tracking with sensor fusion.

• Grant number: 15/26293-0
• Support Opportunities: Scholarships in Brazil - Doctorate
• Start date: July 01, 2016
• End date: July 31, 2019
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal Investigator: Denis Fernando Wolf
• Grantee: Thomio Watanabe
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil

Abstract

Autonomous vehicles development are still a challenge. Despite the latest advances in computer vision the provided solutions are not reliable enough to allow the commercial application of the technology. The high amount of elements and events found in road traffic demand the use of robust solutions that are applicable to any situation. While a completely autonomous vehicle is infeasible the automation is being implemented gradually where automobile manufacturers add new functionalities to ease driving and increase security.This PhD proposal will focus on multiple vehicles tracking in urban traffic through video cameras and LIDAR sensor fusion. Tracking may be defined as the object trajectory estimation. The information about obstacles trajectory helps improve autonomous vehicles navigation as it is the base to analyze other vehicles behavior. In turn, objects tracking depends on objects detection. While obstacles detection have reliable solutions, obstacles tracking based in purely computer vision solutions still present major issues. In some proposed solutions besides the high accuracy error, partial occlusion of obstacles prevents he algorithms to work properly. Computer vision tracking solutions have the three main steps: image acquisition, preprocessing and detection, creating a long pipeline where errors propagate in each one of them. The LIDAR sensor provides straight information about the position of the surrounding objects using point clouds. As there is no need to calculate obstacles distance or velocity using images, it is expected to find more reliable results and to reduce the computational cost in tracking obstacles.For obstacles detection is suggested the use of HOG descriptor and SVMs classifiers. The tracking should be based in the Kalman filter and KLD Adaptive Particle Filter Tracker. The main scientific contribution of this work is the image and point cloud fusion to solve problems in multiple obstacles tracking with partial occlusion in real-time applications.

Scientific publications

(4)

(References retrieved automatically from Web of Science and SciELO through information on FAPESP grants and their corresponding numbers as mentioned in the publications by the authors)

WATANABE, THOMIO; WOLF, DENIS; IEEE. Distance to Center of Mass Encoding for Instance Segmentation. 2018 21ST INTERNATIONAL CONFERENCE ON INTELLIGENT TRANSPORTATION SYSTEMS (ITSC), v. N/A, p. 7-pg., 2018-01-01. (15/26293-0)

WATANABE, THOMIO; WOLF, DENIS F.. Image classification in frequency domain with 2SReLU: A second harmonics superposition activation function. APPLIED SOFT COMPUTING, v. 112, NOV 2021. (15/26293-0)

WATANABE, THOMIO; WOLF, DENIS; DONASCIMENTO, TP; COLOMBINI, EL; DEBRITO, AV; GARCIA, LTD; SA, STD; GONCALVES, LMG. Verisimilar Percept Sequences Tests for Autonomous Driving Intelligent Agent Assessment. 15TH LATIN AMERICAN ROBOTICS SYMPOSIUM 6TH BRAZILIAN ROBOTICS SYMPOSIUM 9TH WORKSHOP ON ROBOTICS IN EDUCATION (LARS/SBR/WRE 2018), v. N/A, p. 6-pg., 2018-01-01. (15/26293-0)

WATANABE, THOMIO; WOLF, DENIS F.; IEEE. Instance Segmentation as Image Segmentation Annotation. 2019 30TH IEEE INTELLIGENT VEHICLES SYMPOSIUM (IV19), v. N/A, p. 6-pg., 2019-01-01. (15/26293-0)