Over the last decade researchers have made serious attempts to approach the Autonomous Driving (AD) vision. In the early stages of these attempts many machine learning
and deep learning algorithms have been utilized without noteworthy success. Unfortunately, their slow detection speed was limiting their potential in driving conditions.
Lately, some new papers were published proposing Convolutional Neural Network
(CNN) based models, designed specifically for real-time object detection, most of them
only trained and tested in general content datasets though. This research was aimed at
exploring in detail three such systems (SqueezeDet, Yolo version 2, Yolo version 3)
and testing their capability in the driving scene. We analyzed their performance on various cases using the KITTI 2D object detection dataset, one of the most representative
datasets for autonomous driving. Besides KITTI, which was our main training and testing dataset, ImageNet and Pascal VOC 2007/2012 have been also used for the pretraining stage of the models.
Collections
Show Collections
