Custom cover image
Custom cover image

Unsupervised Learning of Street Traffic Patterns for Anomaly Detection (PhD Thesis)

By: Material type: TextTextLanguage: English Publication details: Karachi : NED University of Engineering and Technology Department of Computer Science and Information Technology, 2019Description: XVII, 100 p. : illSubject(s): DDC classification:
  • 004.6378242 UME
Summary: Abstract : Anomaly detection through analysis of videos is an important application of intelligent traffic surveillance systems. These systems require the exploration of large amount of traffic data for the detection of normal and anomalous patterns. The developments in computer vision made improvements in video based traffic surveillance system however, it still faced some challenges. Most of the related previous work is based on training data set and require high processing time. Furthermore, cameras that are used in traffic surveillance have high frame rates due to which computation of features at every frame is expensive. In this thesis an unsupervised framework for anomalies detection of street traffic is proposed to minimize the above mentioned requirements. The framework is based on Window frame Based Features Anomaly Detection (WBFAD) algorithm. The algorithm designed to identify street traffic anomalies without the need of labelled training data with reduced processing time. Moreover, the proposed approach utilizes non-trajectory information which is necessary for analysis of activities to give response to situations as they occurred. Our devised approach performs four operations simultaneously; these are i) Computation of appearance, motion and proximity features of multiple moving objects. ii) Estimation of Behaviour Features Variation (BFV). iii) Calculation of proximity matrix by Euclidean distance. iv) Recognition of objects performing anomalous behaviour and generate alarm where necessary. The objects involved in anomalous behaviour are recorded. Displacement of moving objects is determined by directional plot and normal and anomalous patterns are decided by implementation of clustering techniques. Variation in height, area, and angle features are clustered through Density-based Spatial Clustering of Applications with Noise (DBSCAN) and data of proximity matrix is utilized in hierarchical clustering. It is confirmed from experimental results that the proposed framework is capable of providing an efficient and effective solution for anomalies detection.
Holdings
Item type Current library Shelving location Call number Status Date due Barcode
Reference Collection Reference Collection Government Document Section Govt Publication Section 004.6378242 UME Available 96717
Reference Collection Reference Collection Government Document Section Govt Publication Section 004.6378242 UME Available 96718

Abstract :

Anomaly detection through analysis of videos is an important application of intelligent traffic surveillance systems. These systems require the exploration of large amount of traffic data for the detection of normal and anomalous patterns. The developments in computer vision made improvements in video based traffic surveillance system however, it still faced some challenges. Most of the related previous work is based on training data set and require high processing time. Furthermore, cameras that are used in traffic surveillance have high frame rates due to which computation of features at every frame is expensive.
In this thesis an unsupervised framework for anomalies detection of street traffic is proposed to minimize the above mentioned requirements. The framework is based on Window frame Based Features Anomaly Detection (WBFAD) algorithm. The algorithm designed to identify street traffic anomalies without the need of labelled training data with reduced processing time. Moreover, the proposed approach utilizes non-trajectory information which is necessary for analysis of activities to give response to situations as they occurred. Our devised approach performs four operations simultaneously; these are i) Computation of appearance, motion and proximity features of multiple moving objects. ii) Estimation of Behaviour Features Variation (BFV). iii) Calculation of proximity matrix by Euclidean distance. iv) Recognition of objects performing anomalous behaviour and generate alarm where necessary. The objects involved in anomalous behaviour are recorded. Displacement of moving objects is determined by directional plot and normal and anomalous patterns are decided by implementation of clustering techniques. Variation in height, area, and angle features are clustered through Density-based Spatial Clustering of Applications with Noise (DBSCAN) and data of proximity matrix is utilized in hierarchical clustering. It is confirmed from experimental results that the proposed framework is capable of providing an efficient and effective solution for anomalies detection.