A New Approach To Estimate Depth Of Cars Using A Monocular Image

Predicting scene depth from RGB images is a challenging task. Since the cameras are the most available, least restrictive and cheapest source of information for autonomous vehicles; in this work, a monocular image has been used as the only source of data to estimate the depth of the car within the frontal view. In addition to the detection of cars in the frontal image; a convolutional neural network (CNN) has been trained to detect and localize the lights corresponding to each car. This approach is less sensitive to errors due to the disposition of bounding boxes. An enhancement on the COCO dataset has also been provided by adding the car lights labels. Simulation results show that the proposed approach outperforms those who only use the height and width of bounding boxes to estimate the depth.