Self-Supervised Blind Spot Detection System based on Image Processing by Deep Neural Network

In this paper, we propose a single camera-based driver assistance system for blind spots that is selfsupervised and learn without labeled data. The proposed system is developed based on deep neural network and uses gray scale image and optical flow as input. The strength of this paper compared to similar articles is the processing of image sequence information to assess the risk of accident due to moving objects in blind spots. In this paper, based on detection and tracking of the required objects, the existence of vehicles and its relative speed are estimated and the accident can be predicted for the blind spots of the car. The fusion of image with optical flow and features extracted using a deep neural network has increased robustness of the proposed system. In the proposed application of the article, the camera is mounted on the side mirror of the vehicle and is estimated to have a 60% accuracy, the risk of accidentally turning on the sides or changing lanes.