Gordeev, Klyachin "Determination of the Spatial Position of Cars ..."
Purpose: The purpose of this article is to develop a method for calculating the orientation of a car in a photograph from a DVR. This problem arises when designing autonomous vehicles in an urban environment.
Design/Methodology/Approach: The solution to this problem is based on the use of a neural network, the training of which is based on data consisting of 4262 pre-marked photographs provided by the Kaggle community.
Originality/Value: For three angle coordinates regression we employ the VGGnet with all fully-connected layers removed and the last MaxPooling2D output inputed to three parallel regression branch group each contained two branches for computing the confidence for each angle bin and fully regressed angular correction. For three angular coordinates regression we utilize VGGnet16 with all FC layers removed, and the last MaxPooling2D output layer is fedded to inputs of the three parallel regressive groups contained two branches for appropriate angle bin confidence estimation (softmax classification branch) and also fully regressive additional cos(∆θ) and sin(∆θ) angular corrections computing relative to certain bin. Since the current task didn’t require direct 3D car boxes location estimation on the road, the structure of our neural network didn’t have a regression block on spatial dimensions of 3d bounding box for a car, used for similar algorithm testing on such datasets as KITTI и Pascal3D+.
Gordeev A.Y., Klyachin V.A. (2021) Determination of the Spatial Position of Cars on the Road Using Data from a Camera or DVR. In: Popkova E.G., Sergi B.S. (eds) "Smart Technologies" for Society, State and Economy. ISC 2020. Lecture Notes in Networks and Systems, vol 155. Springer, Cham. https://doi.org/10.1007/978-3-030-59126-7_20.