Real-Time Car Detection-Based Depth Estimation Using Mono Camera

Abstract

Object depth estimation is the cornerstone of many visual analytics systems. In recent years there is a considerable progress has been made in this area, while robust, efficient, and precise depth estimation in the real-world video remains a challenge. The approach utilized in this presented paper is to estimate the distance of surrounding cars using a mono camera. Using YOLO (You Only Look Once) in the detection process, by generating a boundary box surrounding the object, then an inversion proportional correlation between the distance and the boundary box's dimensions (height, width) is ascertained. Getting the exact equation between the studied variables; the dependent variables are the distance, and independent variable is the height and width of YOLO boundary box. In the regression model, multiple regression techniques were acclimated to evade heteroskedasticity and multi-collinearity problems. Achieving a real-time detection with a 23 FPS (Frame Per Second) and depth estimation accuracy 80.4%.