Loading...

Traffic Videos Content Analysis Based on Topic Models

Ahmadi, Parvin | 2017

669 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 49778 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Tabandeh, Mahmoud; Gholampour, Iman
  7. Abstract:
  8. Motion pattern analysis in traffic videos can be directly employed for scene analysis, rule mining, abnormal event detection, traffic phase detection, etc. The most successful and newest methods for complex traffic scene analysis are based on topic models. Topic models have been developed for text processing and as they have not been yet optimized for traffic video processing, they are still far away from optimal efficiency. In this thesis, firstly we propose two unsupervised methods for traffic video analysis based on non-probabilistic topic models. In the first proposed method, we use Group Sparse Topical Coding (GSTC) and an improved version of it for learning typical motion patterns occurring in traffic videos. In the second proposed method, by taking the arrangement of optical flow words and also the correlation between the video documents into account, we propose a Sparse Topical Coding (STC) method, called TRSTC for discovering motion patterns in traffic videos. Then, we propose a new dual-layer framework for describing traffic motion patterns based on topic models in a more efficient way. This framework forces the involved topic models to learn the pre-known visually meaningful motion patterns in traffic scenes. Finally, we propose a supervised method which incorporates side information on traffic phases to analyze the traffic videos more efficiently. In all methods, the optical flow features are employed for primary video description. Experimental results on the complex and crowded traffic scenes show that the two unsupervised proposed methods, the proposed framework and the unsupervised proposed method provide more intuitive topics for describing the traffic flows and their results in detection and localization of abnormal events are promising. In comparison with other topic model based methods, according to the values of area under Receiver Operating Characteristic curve (ROC), the first proposed method achieves at least 14% improvement and the second proposed method achieves from 4% to more than 12% improvement in abnormal event detection. Moreover, compared to the conventional single-layer topic model-based frameworks, our proposed framework achieves from 4% to more than 11% improvement in abnormal event detection, in terms of area under ROC, and more than 4% improvement in traffic phase detection accuracy. Furthermore, supervised topic models achieve about 4% improvement in abnormal event detection, compared to the unsupervised ones, in terms of area under ROC
  9. Keywords:
  10. Topic Model ; Abnormal Event Detection ; Traffic Videos ; Motion Patterns ; Traffic Phase Detection

 Digital Object List

 Bookmark

No TOC