Deep Forward Collision Detector in Autonomous Driving Vehicles 2020-01-0138
Forward collision is one of the most challenging concern in the safety of autonomous vehicle. Cooperation between many sensors such as LIDAR, Radar and camera helps to enhance the safety. Apart from the significance of being aware of objects on the drivable area, making an apt decision in the moment is noticeable. In this study, we concentrate on detecting front vehicle of autonomous car using a sensor fusion method, beyond only a detection method. In fact, we devise a solution which provides forward collision warning signal by discriminating between the vehicles moving in and opposite direction of autonomous vehicle, without lane check. Then, the result of classification is combined by the speed of autonomous vehicle as well as the size of detected front vehicle in the images. As a sensor fusion method, this data is utilized to determine whether the front detected car is an obstacle with a potential collision hazard or not. For this reason, we implement a deep neural network with two main parts. The first part is a faster regional convolutional neural network to classify the front and back view of the intended car in conjunction with determining the annotation data. The second part comprises two fully-connected layers with a feedback connection from output to the input for the purpose of emitting hazard signal. This part is a light weight recurrent neural network which distinguishes hazard situation using the result of classification coming from the first part of network along with autonomous vehicle speed, current and past size of front vehicle annotation size. According to experiments, this network was successfully able to specify whether intended front vehicle has the chance of collision or not. This was achieved by a binary classification with respect to the proximity between autonomous vehicle and its front car.