TY - GEN
T1 - Video Footage Highlight Detection in Formula 1 Through Vehicle Recognition with Faster R-CNN Trained on Game Footage
AU - Spijkerman, Ruan
AU - van der Haar, Dustin
N1 - Publisher Copyright:
© 2020, Springer Nature Switzerland AG.
PY - 2020
Y1 - 2020
N2 - Formula One, and its accompanying e-sports series, provides viewers with a large selection of camera angles, with the onboard cameras oftentimes providing the most exciting view of events. Through the implementation of three object detection pipelines, namely Haar cascades, Histogram of Oriented Gradient features with a Support Vector Machine, and a Faster Region-based Convolutional Neural Network (Faster R-CNN), we analyse their ability to detect the cars in real-life and virtual onboard footage using training images taken from the official F1 2019 video game. The results of this research concluded that Faster R-CNNs would be best suited for accurate detection of vehicles to identify events such as crashes occurring in real-time. This finding is evident through the precision and recall scores of 97% and 99%, respectively. The speed of detection when using a Haar cascade also makes it an attractive choice in scenarios where precise detection is not important. The Haar cascade achieved the lowest detection time of only 0.14 s per image at the cost of precision (71%). The implementation of HOG features classifier using an SVM was unsuccessful with regards to detection and speed, which took up to 17 s to classify an image. Both the Haar cascade and HOG feature models improved their performance when tested on real-life images (76% and 67% respectively), while the Faster R-CNN showed a slight drop in terms of precision (93%).
AB - Formula One, and its accompanying e-sports series, provides viewers with a large selection of camera angles, with the onboard cameras oftentimes providing the most exciting view of events. Through the implementation of three object detection pipelines, namely Haar cascades, Histogram of Oriented Gradient features with a Support Vector Machine, and a Faster Region-based Convolutional Neural Network (Faster R-CNN), we analyse their ability to detect the cars in real-life and virtual onboard footage using training images taken from the official F1 2019 video game. The results of this research concluded that Faster R-CNNs would be best suited for accurate detection of vehicles to identify events such as crashes occurring in real-time. This finding is evident through the precision and recall scores of 97% and 99%, respectively. The speed of detection when using a Haar cascade also makes it an attractive choice in scenarios where precise detection is not important. The Haar cascade achieved the lowest detection time of only 0.14 s per image at the cost of precision (71%). The implementation of HOG features classifier using an SVM was unsuccessful with regards to detection and speed, which took up to 17 s to classify an image. Both the Haar cascade and HOG feature models improved their performance when tested on real-life images (76% and 67% respectively), while the Faster R-CNN showed a slight drop in terms of precision (93%).
KW - Faster R-CNN
KW - Haar cascade
KW - Histogram of oriented gradients
KW - Object detection
UR - http://www.scopus.com/inward/record.url?scp=85091335227&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-59006-2_16
DO - 10.1007/978-3-030-59006-2_16
M3 - Conference contribution
AN - SCOPUS:85091335227
SN - 9783030590055
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 176
EP - 187
BT - Computer Vision and Graphics - International Conference, ICCVG 2020, Proceedings
A2 - Chmielewski, Leszek J.
A2 - Kozera, Ryszard
A2 - Orlowski, Arkadiusz
PB - Springer Science and Business Media Deutschland GmbH
T2 - International Conference on Computer Vision and Graphics, ICCVG 2020
Y2 - 14 September 2020 through 16 September 2020
ER -