Relevant Traffic Light Recognition with Deep Learning Approaches

Name: RAFAEL HORIMOTO DE FREITAS

Publication date: 22/10/2019
Advisor:

Namesort descending Role
THIAGO OLIVEIRA DOS SANTOS Advisor *

Examining board:

Namesort descending Role
CLAUDINE SANTOS BADUE Internal Examiner *
PATRICK MARQUES CIARELLI External Examiner *
THIAGO OLIVEIRA DOS SANTOS Advisor *

Summary: Self-driving cars have the important task of recognizing the state (e.g., red, green, or yellow) of
the traffic lights that are relevant, i.e., that define guidance to the car. Common approaches consist
of using the image captured from a forward-looking camera to detect traffic lights in the scene
and classify the respective traffic lights’ states. These approaches have two main limitations: (i)
besides computationally time-consuming, detection usually requires expensive annotations, such
as target objects’ bounding boxes; and (ii) there is still need for a decision-making process in
which relevant traffic lights should be distinguished from the others. This work address these
limitations by investigating two deep learning-based approaches to recognize the relevant traffic
lights’ state: direct-classification and detection-with-classification. In the first, both limitations
are addressed by training the system to direct classify the state of the relevant traffic lights in
the image. In the second, the state recognition is accomplished by detecting traffic lights in
the image with their respective states classified; then the second limitation is addressed with
different heuristics to select a relevant exemplar. Also, a deep regression system with an novel
outliers resilient loss is proposed to predict the coordinates of a relevant traffic light in the
image plane, such that one of the heuristics consists in selecting the closest detection to these
coordinates. Both approaches were evaluated with different real-world datasets. The overall
conclusions are that the direct-classification approach can achieve comparable performance
to detection-with-classification with higher number of easily annotated training images; and
that simple rule-based heuristics have comparable results to the regression system’s heuristic.
Additionally, qualitative assessment with challenging instances revealed both approaches have
similar performance level on grasping the contextual information required to infer the relevant
traffic light. The regression system is also evaluated alone. The results are promising and indicates
that the predicted coordinates can also be used to assist a cheaper classifier to work on a region
of interest.

Access to document

Acesso à informação
Transparência Pública

© 2013 Universidade Federal do Espírito Santo. Todos os direitos reservados.
Av. Fernando Ferrari, 514 - Goiabeiras, Vitória - ES | CEP 29075-910