Search

These last years have been an increasing need regarding the automatic sport analysis. Being soccer one of the most watched sports around the world, action spotting for soccer videos has become one of the main field studies in computer vision. In this project, we have done a deep research and analysis of the State of the Art regarding human recognition and action recognition in soccer videos. Apart from this more theoretical part, we have also executed and compared two of the most pioneering models in sports action spotting today: CALF[3] and NetVLAD++[4] models. In addition, we have been able to make a modification in one hyperparameter, the window size, to evaluate its effect on performance. Finally, we have concluded that we get the highest mAP performance with the NetVLAD++[4] model and that modifying the window size worsens the overall performance in both models, but the individual class performance can benefit from it as the performance is improved in some classes for different window sizes.

This project has been devoted to (i) learning what Multiple Object Tracking (MOT) is, (ii) learning Python, one of the most used languages in Machine Learning and computer vision, and (iii) to evaluate a tracker (TrajTrack), currently being developed at the image processing group (GPI), against the UA-DETRAC dataset. The work has been divided in two parts. On the one hand, we have studied MOT and its main challenges, such as occlusions or identity switches, in order to follow multiple objects throughout a video sequence. To fully understand this problem, we have developed a multiple tennis ball tracker in Python from scratch. On the other hand, we have used TrajTrack, which is evaluated on a pedestrian dataset (MOT17), and adapted it to be evaluated against a car dataset (UA-DETRAC). For this, we have retrained the detection and re-identification models. We have obtained a 98.6% MOTA score for training and a 74.7% MOTA score for testing. These results are comparable with the state-of-the-art techniques., Este proyecto se ha basado en (i) aprender que es el Multiple Object Tracking (MOT), (ii) aprender Python, uno de los lenguajes más utilizados en Machine Learning y en visión por computación, y (iii) evaluar un tracker (TrajTrack), actualmente en desarrollo en el grupo de procesamiento de imagen (GPI), en contra del dataset UA-DETRAC. El trabajo se ha dividido en dos partes. Por un lado, hemos estudiado el MOT y sus principales desafíos, como las oclusiones y los cambios de identidad, a fin de seguir varios objetos a lo largo de una secuencia de video. Para comprender plenamente este problema, hemos desarrollado desde cero un tracker de múltiples pelotas de tenis con Python. Por otro lado, hemos utilizado el TrajTrack, evaluado en un dataset de peatones (MOT17) y lo hemos adaptado para ser evaluado con un dataset de coches (UA-DETRAC). Para ello, hemos reentrenado los modelos de detección y re-identificación. Hemos obtenido un 98.6% en el valor MOTA en el entrenamiento y un 74.7% MOTA en el test. Estos resultados son comparables con las técnicas del estado del arte., Aquest projecte s?ha basat en (i) aprendre que és el Multiple Object Tracking (MOT), (ii) aprendre Python, un dels llenguatges més utilitzats en Maching Learning i visió per computador, i (iii) avaluar un tracker (TrajTrack), actualment en desenvolupament en el grup de processament d?imatge (GPI), en contra del dataset UA-DETRAC. El treball s?ha dividit en dues parts. D?una banda hem estudiat el MOT i el seus principals reptes, com ara oclusions o canvis d?identitat, per tal de seguir diversos objectes al llarg d?una seqüència de vídeo. Per comprendre plenament aquest problema, hem desenvolupat des de zero un tracker de múltiples pilotes de tenis amb Python. D?altra banda, hem utilitzat el TrajTrack, avaluat sobre un dataset de vianants (MOT17) i l?hem adaptat per ser avaluat amb un dataset de cotxes (UA-DETRAC). Per a això, hem re-entrenat els models de detecció i re-identificació. Hem obtingut un 98.6% en el valor MOTA a l?entrenament i un 74.7% MOTA al test. Aquests resultats son comparables amb les tècniques de l?estat de l?art.