Inference Of Lateral Distance to Vehicle from Road Lane Marking using Data from Side Mounted Camera

Sammanfattning

AbstractPilot Assist is an Advanced Driver Assistance Systems(ADAS) functionality used in Volvo cars, that helps the driver to keep the vehicle inside a road lane. In this thesis, we propose a method to estimate the lateral distance between the ego vehicle and the road lane, which can be used to evaluate the Pilot Assist functionality in test cars. Images from side-mounted vision cameras are used for this purpose. Ourapproach involves three steps: (a) Detecting the lane marking on the image, (b) Tracking the lane markings and (c) Estimating the distance between the ego vehicle and the detected lane marking. Three alternative approaches are used to locate the lane markings: to locate the lane as an object, the object detection algorithm You Only Look Once(yolo) or semantically segment the lane using an approach based on Conditional Generative Adversarial Network(cGAN) and a Semantic Segmentation network based on UNet architecture. As the second step, positions of the lane markings are tracked using a Linear Kalman filter in order to estimate their position when the lane markings are missing in certain images in the sequence. In the final step, a look-up table and a neural network approach is used to estimate the distance between the ego vehicle and the lane marking by relating image coordinates to real world distances. A data set is created to train and evaluate the proposed method by labelling images, which is then extended using various data augmentation strategies. The performance of lane detection algorithms are evaluated using Dice coefficient(F1-score), where it was 0.96 for segmentation approach and 0.88 for detection approach on the validation data set. The performance of the Distance estimation software pipeline is evaluated on the test data set, and the average lateral distance error relative to ground truth is found to be approximately 28mm and 32mm when using the segmentation and detection approaches, respectively.