Congestion Detection Through Velocity Estimation Using a Monocular Camera

Abstract

This thesis report aims to answer the following research question: “Is it possible to estimate relative velocities of vehicles surrounding the ego vehicle using a monocular camera with such an accuracy that meaningful conclusions can be made about the current traffic state?” To answer this question, a velocity estimation algorithm is developed in three major parts: object detection, object tracking with detections and velocity estimation using tracked 2D objects. For the detection part, a version of the YOLOv3 (You Only Look Once version 3) single shot detection neural network is used. For object tracking with detections, the Simple Online and Realtime Tracking (SORT) algorithm is used. The last part, velocity estimation using tracked 2D objects, a state-of-the-art method using a neural network is compared to a novel proposed method, using a 2D to 3D map in combination with a kalman filter using a constant velocity model. The results of the detection and tracking parts were good enough to reason that they are used as a base of the velocity estimation algorithm. When comparing the-state-of-the-art velocity estimation algorithm and the novel approach, the errors of the novel approach were significantly higher, and the results of the state-of-the-art methods could not be replicated. This means that the research question of this thesis can be answered with yes, it is possible to estimate relative velocities of surrounding vehicles, however the resulting estimation errors are too high to make meaningful conclusions about the current traffic state.