Autonomous Slalom Maneuver Based on Expert Drivers’ Behavior Using Convolutional Neural Network

Abstract

Lane changing and obstacle avoidance are one of the most important tasks in automated cars. To date, many algorithms have been suggested that are generally based on path trajectory or reinforcement learning approaches. Although these methods have been efficient, they are not able to accurately imitate a smooth path traveled by an expert driver. In this paper, a method is presented to mimic drivers' behavior using a convolutional neural network (CNN). First, seven features are extracted from a dataset gathered from four expert drivers in a driving simulator. Then, these features are converted from 1D arrays to 2D arrays and injected into a CNN. The CNN model computes the desired steering wheel angle and sends it to an adaptive PD controller. Finally, the control unit applies proper torque to the steering wheel. Results show that the CNN model can mimic the drivers' behavior with an R 2 -squared of 0.83. Also, the performance of the presented method was evaluated in the driving simulator for 17 trials, which avoided all traffic cones successfully. In some trials, the presented method performed a smoother maneuver compared to the expert drivers.

SwapTransformer: Highway Overtaking Tactical Planner Model via Imitation Learning on OSHA Dataset

“…Imitation learning is one of the earliest and most successful methods for Autonomous Driving tasks. Authors in [16], [17], and [18] have used different variations of Imitation Learning to solve overtaking or lane changing specifically.…”

“…Authors in [19] showed that using a long short-term memory (LSTM) module for Lane Change (LC) prediction achieves a significant improvement in both adaptive cruise control (ACC) and cooperative adaptive cruise control. In another work [18], the authors used a convolutional neural network (CNN) to show that this network is capable of learning the expert policy excellently.…”

“…During the initial stages of model design and training, hyperparameter tuning was performed on the naive transformer without auxiliary loss or the swapping feature to optimize performance. Various versions of EfficientNet [59] (b2 and b6) and ResNet (18,50) were considered, along with a custom vision encoder. ResNet18 was chosen for its similar performance to other models with lower computational costs.…”