Key points are not available for this paper at this time.
Distance estimation from surrounding vehicles and roadside obstacles plays a crucial part and forms the core of various driver alerts in advanced driver assistance systems (ADAS) in life-threatening situations on the road. Unfortunately, given the high price point at which ADAS components are built, primarily due to the expensive radars and Lidars, the benefits of these systems remain exclusive to specific car segments. The present study attempts to democratize ADAS usage in the vehicle industry by adopting the vision-only approach, thereby making it affordable. The work focuses on building and evaluating algorithms for vehicle detection, classification, and distance estimation using a monocular camera mounted on a vehicle. Two methods have been developed and evaluated: (a) camera-optics-based and (b) image-based deep-learning depth estimation model. The evaluation metrics show promising results for the proposed image-based depth estimation with error correction model, which has the least RMSE and variation in the residuals.
Thombre et al. (Fri,) studied this question.