On the Safety of Autonomous Driving: A Dynamic Deep Object Detection Approach 2019-01-1044

To improve the safety of automated driving, the paramount target of this intelligent system to detect and segment the obstacle such as car and pedestrian, precisely. Object detection in self-driving vehicle has chiefly accomplished by making decision and detecting objects through each frame of video. However, there are diverse group of methods in both machine learning and machine vision to improve the performance of system. It is significant to factor in the function of the time in the detection phase. In other word, considering the inputs of system which, have been emitted from eclectic type of sensors such as camera, radar, and lidar, as time-varying signals, can be helpful to engross ‘time’ as a fundamental feature in modeling for forecasting the object, while car is moving on the way. In this paper, we focus on eliciting a model through the time to increase the accuracy of object detection in self-driving vehicles. In fact, we designed a deep recurrent neural network, which is fed by the output of sparse convolutional neural network as the image feature extractor system. In addition, we modified the energy function of sparse coding to conform the feature vectors in the dynamic environment. Eventually, the proposed system could make a decision about each object not only by utilizing the data during a predefined span of the video, but also enhance the performance of system to determine the object as whether a car or not. Results shows that the performance of system compared to convolutional neural network is more precise in the case of discerning cars.