Auflistung nach Schlagwort "object detection"
1 - 7 von 7
Treffer pro Seite
Sortieroptionen
- KonferenzbeitragA comparative study of RGB and multispectral imaging for weed detection in precision agriculture(44. GIL - Jahrestagung, Biodiversität fördern durch digitale Landwirtschaft, 2024) Benedikt Fischer, Pascal GauweilerPrecision agriculture and specifically mechanical weed control systems have the potential to positively impact our environment by reducing the use of herbicides. In recent years, multispectral cameras have become more and more accessible, which raises the question whether the additional costs of such cameras are worth the potential benefits. In this study, we recorded and annotated a multispectral instance segmentation dataset for sugar beet crop and weed detection. We trained Mask-RCNN models on the RGB and multispectral data in a transfer learning approach and extensively evaluated and compared the results for different scenarios. We found that the multispectral data can improve the weed detection performance significantly in many cases.
- KonferenzbeitragDeep Learning-based UAV-assisted grassland monitoring to facilitate Eco-scheme 5 realization(44. GIL - Jahrestagung, Biodiversität fördern durch digitale Landwirtschaft, 2024) Basavegowda, Deepak H.; Höhne, Marina M.-C.; Weltzien, CorneliaEco-scheme 5 has been introduced to promote biodiversity in permanent grasslands through sustainable land management. While this scheme motivates farmers through result-based remuneration, it also entails a significant monitoring cost in terms of time and money to identify indicators manually. To overcome this burden and facilitate the realization of Eco-scheme 5, we developed an object detection model based on Deep Learning (DL) to automate the indicator species identification. First, we trained and evaluated the model on high-resolution Unmanned Aerial Vehicle (UAV) data. The model achieved an Average Precision (AP) rate of 80.8 AP50, but limited training data and the class imbalance problem among indicators affected the model performance. To address these problems, we enriched training data with proximal images of indicators, resulting in a performance gain from 80.8 AP50 to 95.3 AP50. Our results demonstrate the potential of DL and UAV applications in assisting result-based agri-environmental schemes (AES) such as Eco-scheme 5.
- ZeitschriftenartikelDesPat: Smartphone-Based Object Detection for Citizen Science and Urban Surveys(i-com: Vol. 20, No. 2, 2021) Getschmann, Christopher; Echtler, FlorianData acquisition is a central task in research and one of the largest opportunities for citizen science. Especially in urban surveys investigating traffic and people flows, extensive manual labor is required, occasionally augmented by smartphones. We present DesPat, an app designed to turn a wide range of low-cost Android phones into a privacy-respecting camera-based pedestrian tracking tool to automatize data collection. This data can then be used to analyze pedestrian traffic patterns in general, and identify crowd hotspots and bottlenecks, which are particularly relevant in light of the recent COVID-19 pandemic. All image analysis is done locally on the device through a convolutional neural network, thereby avoiding any privacy concerns or legal issues regarding video surveillance. We show example heatmap visualizations from deployments of our prototype in urban areas and compare performance data for a variety of phones to discuss suitability of on-device object detection for our usecase of pedestrian data collection.
- KonferenzbeitragExtracting Production Style Features of Educational Videos with Deep Learning(Proceedings of DELFI Workshops 2022, 2022) Maya, Fatima; Krieter, Philipp; Wolf, Karsten D.; Breiter, AndreasEnforced by the pandemic, the production of videos in educational settings and their availability on learning platforms allow new forms of video-based learning. This has a strong benefit of covering multiple topics with different design styles and facilitating the learning experience. Consequently, research interest in video-based learning has increased remarkably, with many studies focusing on examining the diverse visual properties of videos and their impact on learner engagement and knowledge gain. However, manually analysing educational videos to collect metadata and to classify videos for quality assessment is a time-consuming activity. In this paper, we address the problem of automatic video feature extraction related to video production design. To this end, we introduce a novel use case for object detection models to recognize the human embodiment and the type of teaching media used in the video. The results obtained on a small-scale custom dataset show the potential of deep learning models for visual video analysis. This will allow for future use in developing an automatic video assessment system to reduce the workload for teachers and researchers.
- KonferenzbeitragIndicator plant species detection in grassland using EfficientDet object detector(42. GIL-Jahrestagung, Künstliche Intelligenz in der Agrar- und Ernährungswirtschaft, 2022) Basavegowda, Deepak Hanike; Mosebach, Paul; Schleip, Inga; Weltzien, CorneliaExtensively used grasslands (meadows and pastures) are ecologically valuable areas in the agricultural landscape and part of the multifunctional agriculture. In Germany, the quality of these grasslands is assessed based on the occurrence of certain plant species known as indicator or character species, with indicators being defined at regional level. Therefore, the recognition of these indicators on a spatial level is a prerequisite for monitoring grassland biodiversity. The identification of indicator species for the status quo of grassland using traditional methods was found to be challenging and tedious. Deep learning-algorithms applied to high-resolution UAV imagery could be the key solution, where UAV with remote sensors can map a large area of grassland in comparison to manual or ground mapping methods and deep learning-algorithms can automate the detection process. In this research work, we use an EfficientDet based algorithm to train an object detection model capable of recognizing indicators on RGB data. The experimental results show that this approach is very promising in contrast to the difficult and time-consuming manual recognition methods. The model was trained with the momentum-SGD optimizer with a momentum value of 0.9 and a learning rate of 0.0001. The model was trained and tested on 1200 images and achieves 45.7 AP (and 85.7 AP50) on test data set. The dataset includes images of four distinct indicator plant species: Armeria maritima, Campanula patula, Cirsium oleraceum, and Daucus carota
- KonferenzbeitragInstance-level augmentation for synthetic agricultural data using depth maps(43. GIL-Jahrestagung, Resiliente Agri-Food-Systeme, 2023) Wübben, Henning; Butz, Raphaela; von Szadkowski, Kai; Barenkamp, MarcoImage augmentation is a key component in computer vision pipelines. Its techniques utilize different levels of data annotation. A lack of methods can be observed when it comes to data that supplies depth maps, in particular synthetic data. We propose a novel augmentation method named DepthAug that utilizes depth annotations in image data and examine its performance in the context of object detection tasks. Results show a boost in MAP score performance compared to previous related methods.
- KonferenzbeitragPollen detection from honey sediments via Region-Based Convolutional Neural Networks(42. GIL-Jahrestagung, Künstliche Intelligenz in der Agrar- und Ernährungswirtschaft, 2022) Viertel, Philipp; Koenig, Matthias; Rexilius, JanThis paper deals with the localization and classification of pollen grains in light-microscopic images from pollen samples and honey sediments. A laboratory analysis of the honey sediment offers precise information of the honey composition. By utilizing state of the art deep neural networks, we show the possibility of automatizing the process of pollen counting and identification. For that purpose, we created and labelled our own data set comprising two pollen classes and trained and evaluated a regional-based neural network. Our results show that the majority of pollen grains are correctly detected. The pollen frequency in the honey sediment is on par with the majority pollen class, however, more samples and further investigation are required to ensure stable results and practicality.