M2BMT: An Automated Ground Truth Generation Algorithm for Lane Detection

Abstract

Safety is one of the primary requirements of automotive manufactures and buyers and regulatory bodies are supporting this by mandating the safety features in vehicles. To achieve safety, multiple sensors such as vision cameras, radars, LIDAR and ultrasound devices are installed in the vehicle at various locations and sensor data is processed continuously using advanced algorithms. Validation of these algorithms is a critical requirement to ensure quality of the system. The present paper proposes an enhancement of the ground truth determination for automated lane detection system. The approach of time-slicing [1] has been built up on the binary framework. But the classical binarization algorithms have some limitations to address the particular domain of lane detection in an unconstrained environment and varied scenarios such as curvy and dashed lane marks. The current paper proposes a novel binarization algorithm based on min-between-max thresholding (MBMT). The adaptive binarization addresses the issue of outlier rejection in an efficient way and handles the effect of shadow, illumination variation and other factors in time-sliced images automatically. In addition the current work identifies limitation of the classical timeslice (TS) based approach even with time MBMT (TMBMT) for ground truth determination and addresses the same through second level of adaptation by spatial MBMT (SMBMT). Finally a complete mathematical model is presented to determine ground truth through the said method of hybrid or Modified MBMT or M 2 BM T . Technology disclosed in this paper is subject matter of pending patent application.

Deep Learning Metadata Fusion for Traffic Light to Lane Assignment

“…The mean accuracy results increase with a larger input image size until they stagnate after 256 px with 85.6 ±0.6 % to 84.2 ±2.5 % for 512 px. The VGG-16 topology achieves the most consistent results among all CNN topologies over all input image sizes with a mean accuracy of 85.2 %.The highest single results are achieved by the AlexNet topology together with an input image size of 256 px (86.4 ±0.5 %) and the VGG-16 topology with an input image size25 The weights are initialised between [0, 1] by using an uncorrelated equal distribution.of 512 px (86.8 ±0.8 %) without a significant difference between both. Moreover, the 256 px input image size produces quite constant simulation results with the lowest 90 % confidence interval values.…”

“…Accuracy was used as a metric for comparison with its 90 % confidence intervals. Each combination was simulated five times with random weight initialisation25 . The mean accuracy results increase with a larger input image size until they stagnate after 256 px with 85.6 ±0.6 % to 84.2 ±2.5 % for 512 px.…”