Autonomous Braking and End to End Learning using Single Shot Detection Model and Convolutional Neural Network

Safety issues concerning autonomous vehicles are becoming increasingly striking. Therefore, taking security issues of autonomous driving into account such as detection and identification of the vehicle in the surrounding is necessary to apply warning messages and braking based on the state of the vehicle. This paper develops an end to end deep learning, using different recognition algorithms, to promote the safety of autonomous vehicles in terms of controlling the steering and speed of a self-driving car. Two convolutional neural network architectures are presented with different number of filters in their layers. The networks were trained to take images as input data and scan the raw pixels and convert them directly into steering angle command and speed value. Also, an object recognition algorithm is provided which detects and determines the objects and their distances from the controlled car to have a collision warning system by using a pre-trained single shot detector model. All predicted speed values and steering angles, alongside the object detection model, are then translated into throttle and braking values while evaluating the models using a simulator and real road videos.