Vehicle trajectory prediction method based on residual attention mechanism

Abstract

The invention discloses a vehicle trajectory prediction method based on a residual attention mechanism. The method comprises the following steps: enabling the trajectories of a target vehicle and surrounding vehicles to serve as model input after the trajectories of the target vehicle and the surrounding vehicles are preprocessed, calculating the weight coefficient of each surrounding vehicle relative to the target vehicle through a residual connection attention module, and reconfiguring an interaction tensor according to the weight coefficient; and combining the new interaction tensor and the coding of the historical trajectory of the target vehicle to extract complete interaction features as the input of a decoder, and finally outputting the probability distribution of a future prediction trajectory through a full connection layer. A mean value of probability distribution is used as an actual track coordinate predicted value, a root mean square error loss value and a back propagation error of the model are calculated, parameters in the model are updated through an Adam optimizer, training is carried out until the loss value of the model is minimum, and good generalization ability is achieved on a verification set and a test set.

Description

technical field [0001] The invention belongs to the field of driving assistance systems, and specifically refers to a vehicle trajectory prediction method based on a residual attention mechanism. Background technique [0002] In recent years, with the public's emphasis on car safety, the application of driving assistance systems in vehicles has become more and more extensive. Among them, the lane departure warning system has been effective in reducing lane departure accidents caused by driving fatigue. However, the existing problems about false alarms of lane departure warnings have always existed. The main problem is to wrongly predict the future trajectory of vehicles. [0003] The current vehicle trajectory prediction technology can be divided into trajectory prediction methods based on dynamic models and trajectory prediction methods based on data learning; methods based on dynamic models include methods based on CV, CTRA, CTRV, etc. Most of the current lane departure wa...

AI Technical Summary

Problems solved by technology

However, the existing problems about false alarms of lane departure warnings have always existed. The main problems focus on wrongly predicting the future trajectory of vehicles. Therefore, improving the accuracy of future vehicle trajectory predictions is directly related to solving the problem of false alarms of lane departure warnings

[0004] The current trajectory prediction method based on data learning not only uses the historical trajectory of the vehicle as the basis for the model to predict the trajectory, but also considers the impact of the interaction between vehicles on the future trajectory of the vehicle. For example, when the driver changes lanes to the left, he will observe and predict the left Whether the driving conditions of the coming vehicle behind the lane will affect the lane change of the own vehicle; construct the interaction tensor according to the road structure and vehicle trajectory, and extract the interaction features in the interaction tensor through the convolutional layer or the fully connected layer; but consider the interaction between vehicles When the trajectory prediction method extracts the interaction features between vehicles, it basically assigns the same weight to each vehicle without bias. In the real driving environment, the driver of the vehicle will only pay attention to some vehicles on the road, and the driver assigns different weights to different vehicles. The degree of attention is different; therefore, there will be a large difference between the unbiased interaction features extracted by assigning the same weight to each vehicle and the real interaction features in the road situation

Images