Person Re-Identification (ReID) is the process of associating images of the same person taken from different angles, cameras and at different times. The task is very challenging as a slight change in the appearance of the person can cause troubles in identifying them. In this thesis, the Re-Identification task is applied in the context of soccer games. In soccer games, the players of the same team wear the same outfit and colors, thus the task of Re-Identification is very hard. To address this problem, a state-of-the-art deep neural network based model named AlignedReID and a variation of it called Vanilla model are explored and compared to a baseline approach based on Euclidean distance in the image space. The AlignedReID model uses two feature extractor branches, one global and one local feature extractor. The Vanilla approach is a variation of the AlignedReID which uses only the global feature extractor branch of the AlignedReID. They are trained using two different loss functions, the Batch Hard and its soft-margin variation. The triplet loss is used, where for each loss calculation a triplet of images is used, an anchor, a positive pair (coming from the same person) and a negative pair. By comparing the metrics used for their evaluation, that is rank-1, rank-5, mean Average Precision (mAP) and the Area Under Curve (AUC), and by statistically comparing their mAPs which is assumed to be the most important metric, the AlignedReID model using the Batch Hard loss function outperforms the rest of the models with a mAP of 81\% and rank-1 \& rank-5 above 98\%. Also, a qualitative evaluation of the best model is presented using Grad-CAM, in order to figure how the model decides which images are similar by investigating in which parts of the images it focuses on to produce their embedding representations. It is observed that the model focuses on some discriminative features, such as face, legs and hands other than clothing color and outfit. The empirical results suggest that the AlignedReid is usable in real world applications, however further research to get a better understanding of the generalization to different cameras, leagues and other factors that may affect appearance would be interesting.

Multiple object tracking (MOT) is the task of detecting multiple objects in a scene and associating detections over time to form tracks. It is essential for many scene understanding tasks like surveillance, robotics and autonomous driving. Nowadays, the dominating tracking pipeline is to first detect all individual objects in a scene followed by a separate data association step, also known as tracking-by-detection. Recently, methods doing simultaneous detection and tracking has emerged, combining the task of detection and tracking into one single framework. In this project, we analyse performance of multiple object tracking algorithms belonging to both tracking categories. The goal is to examine strengths, weaknesses, and real-time capability of different tracking approaches in order to understand their suitability in different applications. Results show that a tracking-by-detection system with Scaled-YOLOv4 and SORT achieves 46.8% accuracy at over 28 frames per second (FPS) on Nvidia GTX 1080. By reducing the input resolution, inference speed is increased to almost 50 FPS, making it well suitable for real-time application. The addition of a deep re-identification CNN reduces the number of identity switches by 47%. However, association speed drops as low as 14 FPS for densely populated scenes. This indicates that re-identification CNNs may be impractical for safety critical applications like autonomous driving, especially in urban environments. Simultaneous detection and tracking results suggests an increased tracking robustness. The removal of a complex data association strategy improves robustness with respect to extended modules like re-identification. This indicates that the inherent simplicity in the simultaneous detection and tracking paradigm can provide robust baseline trackers for a variety of applications. We note that further research is required to strengthen this notion. / Multipel objektspårning handlar om att detektera alla objekt i bilder och associera dem över tid för att bilda spår. Det är ett viktigt ämne inom datorseende för flera applikationer, däribland kameraövervakning, robotik och självkörande fordon. Idag är det dominerande tillvägagångsättet inom objektspårning att först detektera alla objekt och sedan associera dem i ett separat steg, också kallat spårning-genom-detektion. På senare tid har det framkommit nya metoder som detekterar och spårar samtidigt. I detta projekt analyserar vi prestanda av metoder som tillämpar båda tillvägagångssätt. Målet med projektet är att undersöka styrkor, svagheter och hur väl metoderna lämpar sig för att användas i realtid. Detta för att förstå hur olika objektspårare kan anpassas till olika praktiska applikationer. Resultaten visar att ett system som tillämpar spårning-genom-detektion med Scaled-YOLOv4 och SORT, uppnår 46.8% noggrannhet med en hastighet på över 28 bildrutor per sekund. Detta på en Nvidia GTX 1080. Genom att minska bildupplösningen når hastigheten nästan hela vägen upp till 50 bildrutor per sekund, vilket gör systemet väl lämpat för realtidsapplikation. Genom att addera ett djupt nätverk för återidentifiering minskar antalet identitetsbyten med 47%. Samtidigt minskar också hastigheten för spårning till 14 bildrutor per sekund i välbefolkade miljöer. Detta indikerar att djupa nätverk för återidentifiering inte lämpar sig för säkerhetskritiska applikationer såsom självkörande fordon. Särskilt i urbana miljöer. Resultat för system som detekterar och spårar samtidigt antyder att de är mer robusta. Genom att ta bort komplexa strategier för associering blir systemen robusta mot ytterligare moduler såsom återidentifiering. Det ger en indikation på att den inneboende enkelheten i dessa system resulterar i objektspårare som kan fungera som grunder i många olika applikationer. Vi noterar att ytterligare forsking behövs för att styrka denna idé.