Interpolation and simulation of autonomous driving camera data for vehicle position synchronization

Abstract

In order to meet the data requirements of the virtual simulation test of autonomous driving, we use the camera’s single-sample video data, interpolate frames to generate multiple camera simulation data. And then realize the simulation of extended front-end perception function at the data level. This paper proposes a video sampling frame simulation reconstruction mechanism based on sampling vehicle pose information in real scenes. Calculate the target position of the simulated vehicle according to the simulation requirements, and establish a simulation node on the sampling path. Taking the position difference between the simulated node and the real node as the offset, the DAIN algorithm is used to insert the image data into the target node. It is possible to realize simulation data generation of autonomous driving camera with variable vehicle speed/sampling frequency. Combined with the camera’s internal and external parameters, the coordinate system transformation of the annotation results is carried out to realize the inheritance of the annotation results of the simulation data. This paper combines the nuScenes open source database to test the authenticity of the synchronous simulation data results. The results show that the mean SSIM of the synchronous simulation image data and the real data is 0.71684, indicating that the simulation data has high authenticity. Based on yolov4, the perceptual function of the simulated data is verified. The average value of the SSIM of the simulated image data and the real data recognition frame is 0.984504795, and the perceptual recognition results are similar. The synchronous mapping result of the marked 3D-BOX box to the simulation data space is correct. Camera simulation data can well meet the needs of autonomous driving development and testing in perception and recognition. It can provide data support for autonomous driving simulation test.