Detection and tracking of motorcycles in urban environments by using video sequences with high level of oclussion

This thesis presents an investigation into detection, classi_cation and tracking of occluded motorcycles from urban tra_c scenes. The _nal aim is to develop an accurate system that allows automatic detection and tracking of motorcycles, which are the most frequent vulnerable user of urban tra_c in emerging countries. Operators of urban tra_c surveillance system could enhance the monitoring of this users and even prevent the high accidentally rate that they represent. Initially, a Motorcycle classi_er for urban scenarios is implemented using a pre-trained convolutional neural network for feature extraction. Motorcycles and cars are classi_ed by using the extracted features from a CNN network, and classi_ed using an SVM. The strategy is evaluated in an urban tra_c dataset, achieving a 99.4% accuracy working with three classes, and 99.3% accuracy with _ve classes. Given the good classi_cation results, we move to detection and classi_cations of vehicles in a urban dataset. A hybrid strategy, which combines GMM for object detection and use of CNN for feature extraction and posterior classi_cation, is _rst considered. Then, a two stage detector as Faster R-CNN is used for object detection and classi_cation. The pre-trained Faster R-CNN model achieves an F1 score of 68% outperforming the hybrid model, which achieves 58 %. Based in the good results obtained by a two stage detector as Faster R-CNN, we propose EspiNet, which is a more compact network able to detect and classify motorcycles under high occlusion in congested urban tra_c environments. The method detects and classify motorcycles even under camera movements, objects overlapping and stationary objects. Due to the absence of urban annotated motorcycle datasets, we introduce a new dataset of 7500 and 10,000 annotated images, captured under real tra_c scenes, using a drone mounted camera. The proposed model achieves an F1 Score of 95.3% with an AP of 89.32 %. Overcoming the results of state of the art detectors trained end to end in the introduced Urban Motorbike Dataset (UMD). For benchmark proposes, we compare with a single stage detector such as Yolo v3 and two stage detectors as Faster R-CNN (VGG16 based). The proposed model is used to improve tracking, in a Multiple Object Tracking implementation based on a Markov Decision Process, and in a Deep Learning MOT tracking mechanism. The detection results with a high con_dence hypothesis, improve the tracking processes achieving a Multiple Object Tracking Accuracy (MOTA) of 86.1% and 87.6% respectively, overcoming the state of the art results presented in tracking benchmarks as the used in KITTI dataset. The thesis concludes with a critical analysis of the presented work and a general outlook for future research proposes

• Autores: Espinosa Oviedo, Jorge Ernesto
• Tipo de recurso: Doctoral thesis
• Fecha de publicación: 2019
• Institución: Universidad Nacional de Colombia
• Repositorio: Universidad Nacional de Colombia
• Idioma: spa
• OAI Identifier: oai:repositorio.unal.edu.co:unal/76455
• Acceso en línea: https://repositorio.unal.edu.co/handle/unal/76455; http://bdigital.unal.edu.co/72867/
• Palabra clave: Object detection; motorcycle detection; motorcycle tracking; multiple object tracking; object detection under high level of occlusion; object tracking under high level of occlusion; Deep Learning, Convolutional Neural Networks; Faster R-CNN
• Rights: openAccess
• License: Atribución-NoComercial 4.0 Internacional