Prediction of vehicle intentions for advanced autonomous driving

Abstract

Durante los últimos años el interés por los sistemas de predicción avanzada de trayectorias de vehículos y de intenciones ha crecido notablemente. Inicialmente la predicción tanto de trayectorias como de maniobras se ha centrado en observaciones realizadas desde puntos estáticos tales como la infraestructura. Esto se ha debido a la falta de bases de datos adecuadas para la predicción desde un punto de vista centrado en el vehículo.

Esta tesis aborda el problema de la predicción de maniobras y trayectorias en entornos de autopistas con un enfoque basado en aprendizaje máquina. Ante la ausencia de bases de datos apropiadas para su desarrollo se tomó la decisión de realizar una base de datos específica para la predicción tanto de trayectorias como de maniobras. Así nace The PREVENTION dataset.

El sistema de predicción de maniobras se basa en una arquitectura de redes neuronales convolucionales que clasifica una imagen de entrada en tres posibles categorías correspondientes con las acciones de cambio de carril a la izquierda y a la derecha y la acción de continuar en el carril actual.

Para poder comparar el rendimiento del sistema de predicción de intenciones con la capacidad humana de predicción se ha realizado un estudio que evalúa la capacidad de predecir o detectar cambios de carril, así como la tasa de acierto de estos.

El sistema de predicción de trayectorias adapta una red neuronal convolucional desarrollada para clasificación. La red ha sido modificada para tomar a la entrada una imagen 3D que codifica una secuencia de imágenes de un solo canal. La salida es similar a la entrada solo que codifica la misma secuencia en el futuro.

During the last years, the interest in advanced vehicle trajectory and intention prediction systems has grown remarkably. Initially, the prediction of both trajectories and maneuvers has been focused on observations made from static points of view, such as the infrastructure because of the lack of appropriate vehicle-centered datasets. This thesis addresses the problem of predicting maneuvers and trajectories in highway environments with a machine learning approach. A specific database for the prediction of both trajectories and maneuvers was created because of the lack of appropriate ones. Thus, The PREVENTION dataset was born. A database recorded from an onboard perspective that includes almost 6 hours of recordings. It has 2 cameras, a rotating laser, and 3 radars, as well as a differential localization system and an inertial measurement unit. This database includes several manual annotations that allow the identification of vehicles and lane changes as well as the positions of the vehicles. The maneuver prediction system is based on a convolutional neural network architecture that classifies an input image into three possible categories corresponding to left and right lane-change actions and the lane-keeping action. The input image consists of three channels, each one has a specific purpose. The red channel represents the environment, the appearance of the scene. The blue channel is used to select the prediction target, from which the current and past contours are drawn with different intensity levels creating a kind of trail that shows the dynamics of the vehicle. The green channel is used to draw the trails of all the surrounding vehicles that actuate as conditioning elements for the vehicle represented in the blue channel. This representation does not limit the number of vehicles in the scene and the number of samples that can be represented is 255. To compare the performance of the intention prediction system with the human prediction capacity, a study was carried out to evaluate the capacity to predict or detect lane changes, as well as the lane change IV accuracy rate. The metrics used to compare the performance of people and the prediction system are the accuracy rate and the anticipation. The trajectory prediction system adapts a convolutional neural network developed for the classification of cells in clinical images. This network extracts features at different depth levels to finally generate an output image. The network has been modified to take a 3D input image that encodes a single-channel image sequence. The output is similar to the input, but it encodes the same sequence in the future. The vehicles are represented on a bird’s eye view that generates a graphic representation of the scene. Besides, elements such as road markings can be added to this representation. The network learns the underlying mechanics and interactions between vehicles and the environment to generate the positions of those vehicles in the future. A Kalman filter with a constant speed model has been implemented as a baseline to compare with the obtained results. Trajectory prediction has been evaluated using several common metrics in the literature, such as RMSE, MAE, ATE and FTE.