Today, self-driving vehicles are a popular idea. Of course, it is treated with caution; Because the decision-making process in this research process is supposed to be an intelligent tool instead of a human, that's why engineers and industrialists are trying to create self-driving vehicles that can safely support them in the real world. One of the issues in this field is trajectory planning for autonomous vehicles, one of the important parts of which is trajectory prediction. Predicting a vehicle's path is an essential capability for autonomous vehicles navigating complex urban traffic scenes. In the discussion of predicting trajectories for a car, there is an issue of uncertainty, because a driver may make a decision different from our guess at any moment. On this basis, the prediction of multiple trajectories has recently been considered to create all the possibilities for a driving agent. In this research, we use a deep learning mixed model named multiple trajectory Prediction (MTP) model and try to train the model by introducing an approach based on the model's training error function. In this research, focusing on how the driving agent accelerates along a trajectory, a training error function is introduced to improve and improve the quality of the acceleration of the predicted trajectories by training the model. It is expected that in two specific scenarios, one of very high speeds and the other of high accelerations, favorable results will be obtained with this combined error. The application of this training error function, which is introduced as acceleration error (Acc-Loss), has been eva‎luated by combining with the main error of the model (MTP-Loss) in order to train the MTP model. By conducting simulations using the nuScenes real-world dataset and raster images from the top view of driving scenes, we obtained favorable results and showed that the predictions improve in the two mentioned scenarios.

مريم ذكري

سمانه حسيني

فريد شيخ الاسلام , جواد عسگري مارناني