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 Oonvolutional 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-ONNs 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%).
展开▼
机译：一级方程式，其伴随的电子运动系列，提供了具有大量摄像机角度的观众，船上摄像机提供了最令人兴奋的事件观点。通过实施三个物体检测管道，即哈尔级联，带有支撑矢量机的导向梯度特征的直方图，以及基于速度的基于区域的Oonvolutional神经网络（更快的R-CNN），我们分析了他们在真实中检测车辆的能力 - 使用培训图像从官方F1 2019视频游戏中获取的培训图像。该研究的结果得出结论，更快的R-ONNS最适合准确检测车辆，以确定实时发生的崩溃等事件。这一发现通过精度和召回分别分别为97％和99％的精确度。使用HAAR级联时的检测速度也使其在精确检测不重要的情况下具有吸引力的选择。 HAAR级联以精度成本（71％）以每张图像的最低检测时间达到0.14秒。在检测和速度方面，使用SVM的使用SVM实现Hog特征分类器的实现是不成功的，该速度最多需要17秒以对图像进行分类。 Haar Cascade和Hog特征模型在现实寿命图像（分别为76％和67％）时，它们的性能提高了它们的性能，而更快的R-CNN在精度（93％）方面表现出略微下降。
展开▼
