To make a road vehicle drive autonomously without human intervention, the first step is to perceive the road environment and detect road features like lane markings, road signs, and obstacles. Perceiving the road environment is carried out by obtaining the 3D information of the environment with respect to the ego vehicle using perception sensors like a Stereo camera, rotating LiDAR (16/32/64/128-channel), or solid-state LiDAR. However, the major drawback of these sensors is their per-unit cost which is very high compared with the full cost of mid-range passenger cars; thus, preventing the implementation of autonomous driving features in these vehicles that are most widely sold in developing countries like India. Hence, the primary objective of this research work is to design and develop a cost-effective Hybrid Perception System (HPS) that consists: of a four-channel LiDAR (for depth estimation), a Monocular Camera (for obstacle and lane detection), and a specially designed interfacing electronic hardware. The special-purpose electronic hardware is used for driving and synchronizing four-channel LiDAR (laser diode and photodiode). The HPS also has a microcontroller for data acquisition and transferring the data computing system for further processing. The performance of the HPS is being evaluated based on Automated Driving tasks like obstacle detection, and lane detection.
This seminar talk focuses on algorithm development for performing these tasks. In the first part of the seminar, a methodology developed to detect, track, and estimate the velocity and distance travelled by the moving obstacle using a monocular camera by solving the scale ambiguity of a monocular camera will be discussed. A Modified Spatial Calibration (MSC) approach has been developed to compute the Scaling Factor (SF) to convert the pixel-based velocity into physical units. Later, the accuracy of the estimation is improved by using SF approximation models ensuring reliable velocity estimation even in the varying day and night conditions. In the second part of the seminar, a novel dynamic thresholding approach-based lane detection algorithm using the Adaptive Confident Road Region Estimation (ACRRE) that adapts to the varying road surface intensity levels will be discussed. This algorithm has been tested with various datasets like KITTI, tuSimple, and own collected datasets and found that the developed algorithm performs better in the mentioned datasets than the other methods. Further, its efficacy has been tested by using it on the Lane Centering Assist System of the prototype intelligent vehicle; thereby showcasing its practical applicability in real-world scenarios.