Monoscopic Automotive Ego-motion Estimation and Cyclist Tracking

In traffic accidents cyclists are always counted as a vulnerable group suffering from heavy injuries and fatalities. A particularly dangerous type of collision involves trucks that turn to the right without recognizing cyclists driving on their lane in a blind spot. Even though it is not the most frequent type of accident the chance of survival for a cyclist involved is low. One of the main causes for the collision is that the truck drivers only have a limited field of vision. The cyclists in the surroundings are hard to perceive due to their smaller size. Besides, it is difficult to predict their behavior. The velocity of a cyclist is usually comparable to a slowly running car and they must share the same road with other traffic participants, which makes them easy to be occluded by other vehicles. Hence, this reduces the truck driver’s reaction time once they are noticed. This also explains that the heaviest accidents involving trucks and cyclists often happen when a truck turns right at an intersection. In order to solve this problem on an intelligent level we are aiming at developing a driving assistance system for trucks to avoid possible accidents with cyclists. The main task is to detect the cyclists with the help of a state-of-the-art hardware setup consisting of a single-row Light Detection And Ranging(LIDAR)Sensor in combination with a camera. Based on the detection, the movement of the cyclist is estimated and its behavior is predicted so that the risk of accidents can be assessed. An intelligent warning strategy alarms the truck driver in dangerous situations to avoid accidents. The aim of this work is to estimate the metric trajectory of the ego-vehicle and the cyclist using a single camera and a low-cost single-row laser scanner. For estimating the ego-motion without scale we complement existing methods to satisfy the requirements of our application. The main challenge of monoscopic Visual Odometry is the unobservability of the scale of the scene. The focus of this project lies on estimating the scale from scene implicit information combined with a priori knowledge about the scenery. On the other hand external sensors are applied to obtain more data about the scale and find an optimal configuration for estimating a scaled trajectory of a car in various environments.