This paper presents a method to estimate the relative distance and speed of vehicles approaching from the side and rear for a blind spot detection system or lane change assistance. Vehicles with the same relative distance can be assigned different pixel positions on the image frame depending on the road curvature. Thus, based on a lane model and vehicle dynamics, a Perception-Net estimation algorithm has been designed to recognize the rear-side lanes and their curvature under various noisy environments. Vehicles approaching from the rear are detected by calculating the velocity vector of feature flows, which are a type of motion analysis, on the image frame. Through various simulations, it is concluded that the proposed method can provide robust estimations of the relative distance and speed between vehicles with satisfactory error tolerance.