Object Detection in Self Driving Cars Using Deep Learning

Abstract

In the Computer Vision domain, there has been continuous growth and development with main focus so as to facilitate a smooth interaction between Machines and human. Perception, planning and control are the main aspects that make up the Self-driving system. Perception subsystem converts the raw data collected by sensors or other information capturing devices into a model of the environment surrounding us. Planning subsystem analyses this model of the surrounding environment and makes certain purposeful decisions based on the inferences obtained from the analysis. Finally, the Control Subsystem is responsible for execution of the actions or the decisions planned previously. The scope of this project is to study and analyze the problems faced in the Perception subsystem in the domain of detecting objects for autonomous cars. Previously, technologies like Radar, LiDAR, GPS and various other sensors had been employed for Driverless cars for mapping the surroundings of the car. However, in the recent past, some deep neural network (DNN) architectures like YOLO (You Only Look Once), and SSD (Single Shot MultiBox Detector) have been developed which are capable of detecting objects even when live video is considered as the input, thus having potential to be included as a part of the Driverless car systems. Selection of a model having considerable accuracy and producing results at a faster rate is very much essential so as to meet the requirements of object detection in driverless cars. In this project, we have used Caffe, which is developed by Berkeley AI Research and Community contributors as the deep learning framework. Keeping in mind the factors that contribute to the selection of a good model, we have chosen SSD model along-side MobileNet Neural network as the base architecture as it results in both faster rate of result production and has a moderate accuracy.