Currently, the use of Unmanned Aerial Vehicles (UAVs) in natural and complex environments has been increasing, because they are appropriate and affordable solutions to support different tasks such as rescue, forestry, and agriculture by collecting and analyzing high-resolution monocular images. Autonomous navigation at low altitudes is an important area of research, as it would allow monitoring parts of the crop that are occluded by their foliage or by other plants. This task is difficult due to the large number of obstacles that might be encountered in the drone’s path. The generation of high-quality depth maps is an alternative for providing real-time obstacle detection and collision avoidance for autonomous UAVs. In this paper, we present a comparative analysis of four supervised learning deep neural networks and a combination of two for monocular depth map estimation considering images captured at low altitudes in simulated natural environments. Our results show that the Boosting Monocular network is the best performing in terms of depth map accuracy because of its capability to process the same image at different scales to avoid loss of fine details.

Nowadays unmanned aerial vehicles (UAV’s) have an important role in Precision Agriculture, due to this it is necessary the autonomous navigation in crop environments. Motivated by the latter, in this paper we present a review on Deep Learning (DL) techniques as a promising alternative for providing real-time obstacle detection and collision avoidance for autonomous UAVs. Furthermore, this article focuses on enumerating the most open challenges for current DL-UAV solutions to allow the navigation at low altitudes in complex crops.

Presently, depth estimation research has focused on deep learning, mainly on monocular depth estimation, because this technique has proven to be an excellent alternative for methods that use expensive sensors or require high computational consumption, providing higher performance and accuracy. Despite there being a large number of works, most have focused on indoor and urban datasets. Due to this, we review the variants of the residual networks being trained with images of natural environments with a high presence of vegetation. We proposed a new dataset specialized in natural environments using the AirSim simulation to evaluate the variants of the ResNet and ResNeXt families and verify that these can be used to detect leaves and thin branches with their corresponding depth. We conclude that a deeper residual neural network is not always the best option for monocular depth estimation. It is necessary to consider the type of network architecture, computing resources available, and the content and size of the data set. Keywords: Deep Learning, Residual Networks, Monocular Depth Estimation RESUMEN: Actualmente, las investigaciones en estimación de profundidad se han centrado en el uso del aprendizaje profundo, principalmente en estimación de profundidad monocular, debido a que esta técnica ha demostrado ser una excelente alternativa para los métodos que utilizan sensores cotosos o requieren un alto consumo computacional, proporcionando mayor rendimiento y exactitud. A pesar de haber una gran cantidad de trabajos, la mayoría se han centrado en conjuntos de datos de escenas de interiores y urbanas. Debido a esto, en este trabajo se realiza una comparativa de las variantes de redes residuales con imágenes de ambientes naturales con alta presencia de vegetación, obteniendo excelentes resultados. Propusimos un nuevo conjunto de datos especializado en entornos naturales utilizando la simulación AirSim, para evaluar las variantes de las redes residuales y comprobar que estas pueden usarse para detectar hojas y ramas delgadas con su correspondiente profundidad. Llegamos a la conclusión de que una red residual neuronal más profunda no siempre es la mejor opción para la estimación de profundidad monocular. Es necesario considerar el tipo de arquitectura de la red, los recursos informáticos disponibles y el contenido y tamaño del conjunto de datos. Palabras clave: Aprendizaje Profundo, Redes Residuales, Estimación de Profundidad Monocular