Semi-automatic labelling and tracking of targets for autonomous driving

• Main Author: Silva, Nuno Miguel Soares
• Publication Date: 2018
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10773/25902
• topic: Autonomous driving; Autonomous vehicles; ATLASCAR; Data Clustering; Calibration; Data Labelling; Object Detection; Data Fusion; ADAS; LIDAR; Image Processing

In the scope of the ATLASCAR2 project, this dissertation is based on the development (using ROS) of a driving assistance system that implements an interface to detect, track and label targets on the road. The detection and tracking are done using LIDAR sensor data and a camera. Firstly, an algorithm is implemented to be used in the camera based on the appearance of the target to be tracked. Next, a range based algorithm is developed using the data acquired from the sensors to follow objects in a tridimensional space. Finally, because it is possible to detect and track objects using the image and the lasers, the combination of the algorithms is done by projecting what is captured from the sensors in the camera image, being possible to obtain a more accurate and robust tracking. To evaluate the algorithm some datasets were used with the labelled data from the several detected and followed objects. To perform this work the sensors and the camera need to be properly calibrated. To do this, the calibration between the several devices was done using an application that, by passing a ball in front of the sensors, the position values of each sensor relatively to a given reference device are found. Within the calibration, this dissertation also includes an improvement of the ball detection algorithm in the image obtained by the camera.