Semantic Segmentation in Anomaly Detection: Researching the feasibility of utilizing semantic segmentation in anomaly detection
Author
Bidstrup, Michael
Term
4. term
Publication year
2021
Submitted on
2021-06-18
Pages
53
Abstract
Dette projekt undersøger, om semantisk segmentering kan forbedre detektering af anomalier i video. Semantisk segmentering betyder, at hvert pixel i et billede klassificeres (fx vej, bil, person), mens anomalidetektion handler om at finde usædvanlige hændelser. Vi ser også på, hvordan man bedst udnytter datasæt, når der mangler ground truth for et eller begge systemer. Vi analyserer segmenteringsresultater fra modeller på anomalidetektionsdatasættet UCHK Avenue, og semantisk segmenteringsdatasættet Cityscapes annoteres manuelt med rammespecifikke anomalier. Anomalidetektionen udføres af et generative adversarial network (GAN), der forudsiger fremtidige videorammer ud fra rumlige og tidslige mønstre. Vi sammenligner systemets ydeevne med segmenterede rammer som input med ydeevnen ved brug af almindelige farvebilleder (RGB-rammer). Testene viser, at segmenterede rammer kan opnå en nøjagtighed, der er konkurrencedygtig med RGB-rammer, forudsat at segmenteringen er tilstrækkelig præcis.
This project investigates whether semantic segmentation can improve video anomaly detection. Semantic segmentation labels every pixel in an image (for example as road, car, or person), while anomaly detection aims to spot unusual events. We also explore how to make effective use of datasets when ground-truth labels are missing for one or both systems. We analyze segmentation results from models on the anomaly detection dataset UCHK Avenue, and the semantic segmentation dataset Cityscapes is manually annotated with frame-specific anomalies. The anomaly detection system is a generative adversarial network (GAN) that predicts future video frames using spatial and temporal patterns. We compare the system’s performance when using segmented frames as input with its performance on standard color images (RGB frames). Tests show that segmented frames can achieve accuracy competitive with RGB frames, provided the segmentation is sufficiently accurate.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Keywords
semantic ; segmentation ; anomaly ; abnormal ; avenue ; cityscapes
Documents