Browse Publications Technical Papers 2020-01-0138
2020-04-14

A Forward Collision Warning System Using Deep Reinforcement Learning 2020-01-0138

Forward collision warning is one of the most challenging concerns in the safety of autonomous vehicles. A cooperation between many sensors such as LIDAR, Radar and camera helps to enhance the safety. Apart from the importance of having a reliable object detector, the safety system should have requisite capabilities to make reasonable decisions in the moment. In this work, we concentrate on detecting front vehicles of autonomous cars using a monocular camera, beyond only a detection method. In fact, we devise a solution based on a cooperation between a deep object detector and a reinforcement learning method to provide forward collision warning signals. The proposed method models the relation between acceleration, distance and collision point using the area of the bounding box related to the front vehicle. An agent of learning automata as a reinforcement learning method interacts with the environment to learn how to behave in eclectic hazardous situations. The agent follows a deterministic but variable structure learning automata in order to find the collision point in different status. The proposed learning automata method has a nested structure in its states and every state has a memory to place the time duration between the entrance and the collision point. The Agent makes decisions based on a specific number of previous data to engross the impact of time. Since it is almost impossible to capture the data of hazardous situations, we design the essential scenarios via a virtual simulation environment. Eventually, after some trails and errors the algorithm reaches to a convergence point and extracts the model of front vehicle motions.

SAE MOBILUS

Subscribers can view annotate, and download all of SAE's content. Learn More »

Access SAE MOBILUS »

Members save up to 18% off list price.
Login to see discount.
Special Offer: Download multiple Technical Papers each year? TechSelect is a cost-effective subscription option to select and download 12-100 full-text Technical Papers per year. Find more information here.
X