Reactive Collision Avoidance of UAVs with Simple Pin-hole Camera Based Passive Stereovision Sensing

An effective reactive collision avoidance algorithm is presented in this paper for unmanned aerial vehicles (UAVs) using two simple inexpensive pinhole cameras. The vision sensed data, which consists of the azimuth and elevation angles at the two camera positions, is first processed through a Kalman filter formulation to estimate the position and velocity of the obstacle. Once the obstacle position is estimated, the collision cone philosophy is used to predict the collision over a short period of time. In case a collision is predicted, steering guidance commands are issued to the vehicle to steer its velocity vector away using the nonlinear differential geometric guidance. A new cubic spline based post-avoidance merging algorithm is also presented so that the vehicle rejoins the intended global path quickly in a smooth manner after avoiding the obstacle. The overall algorithm has been validated using the point mass model of a prototype UAV with first-order autopilot delay. Both extended Kalman filtering (EKF) and unscented Kalman filtering (UKF) have been experimented. Both are found to be quite effective. However, performance of UKF was found to be better than EKF with minor compromise in computational efficiency and hence it can be a better choice. Note that because of two cameras, stereovision signature gets associated with optical flow signature thereby making the overall signature quite strong for obstacle position estimation. This leads to a good amount of success as compared to the usage of a single pinhole camera, results of which has been published earlier.