Obstacle warning systems

Question

Which is the best algorithm to design an obstacle avoidance system with an underlying predictive logic ?

I am using a system of microwave radars which gives the coordinates and velocities of various obstacles. Based on these inputs an algorithm should identify the most critical obstacles and warn to the driver of car / aircraft. The main point of concern is the prediction part. The system should not only identify present obstacles but also look for future collisions based on the velocity components. Which are the possible strategies to explore (Kalman, extended Kalman, Neural Networks, evolutionary algorithms)

Answer 1

Before you consider which strategies are appropriate, you need to consider what constraints describe the motion of both obstacles and the vehicle, generally those constraints are combined into a set of differential equations. Without this information, it isn't possible to make any reasonable recommendation.

• Kalman filters are an optimal predictor for a linear system, or systems that can be approximated about one or more points with linear systems.
• EKFs or UKFs are more commonly used for non-linear systems.
• Neural Networks and Evolutionary algorithms are probably not worth considering - there are much better solutions for the obstacle tracking problem.

Pure obstacle avoidance goes back to old school robotics with minimal computational resources. This includes Brook's Subsumption architecture and Khatib's Potential Fields. However, most of the effort of collision avoidance systems isn't actually related to avoiding obstacles, but rather on planning efficient paths between configurations (or states). The current state of the art for vehicular planning (as used for the DARPA urban challenge and some JPL rovers), appears to be a hybrid of offline kinodynamic trajectory generation and the D*-Lite algorithm.