Zahra Dastjerdi - Sharif University of Technology,Tehran, Iran
Saeed Bagheri Shouraki - Electrical Engineering Department, Sharif University of Technology, Tehran, Iran
Fateme Azizabadi - Sharif University of Technology,Tehran, Iran

Self-driving car can move like an experienced driver without human intervention. For this purpose, she must be able to fully understand and feel her surroundings like a human being. Accurate understanding of the surrounding environment in real time is one of the important factors that affect the performance of self-driving vehicles. Perception refers to the ability of an autonomous vehicle to collect sensor data, extract relevant knowledge, and develop contextual understanding of the environment, To understand the environment, we focused on the two parts of obstacle detection and drivable area detection in front of the car. For this purpose, we use the Kitti dataset. This dataset is the largest dataset for machine vision algorithms for self -driving cars, which includes images captured by a moving vehicle from various positions in the streets of Karlsruhe city. For the road detection part, we were able to reach a precision value of ۹۷.۲۳% and a processing rate of ۲۲ frames per second by replacing the convolutional layer instead of the fully connected layer and combining the spatial features with the features of the deep layers, which is acceptable accuracy compared to the previous results achieved while having the desired speed. In the obstacle detection section, we first regressed the three-dimensional features of the object using a convolutional neural network, and then by combining these features with the limitations of the two-dimensional frames, we were able to predict the three dimensional frames and direction for the vehicles on the road. Using regression instead of deep convolutional networks greatly improved the speed of model training and testing. Finally by definingthe loss function as a sum of coefficients, we were able to reduce the loss value to - and reach a processing rate of ۲۰ frames per second ۰.۸۰۲۶ and an AP of ۸۳.۱۲%, whichaccording to the results obtained this network has a favorable and acceptable performance compared to the results of previous works.

Autonomous vehicles, Road detection, ۳d detection, Regression