Title Dynamic object detection and recognition in an UAV image sequence

 

Unmanned Aerial Vehicle (UAV) have been used more and more lately and gained popularity throughout both in the general public and the engineering world. Image sequences captured from a camera that is attached on a small drone usually contain background such as tree, construction (building and road), grass, sky. In [Stephan2014] partition each frame in multiple segments and to label them with these basic categories. In this project, we will extent Stephan’s works ([Stephan2014]) to detect and recognize dynamic objects in the captured scene such as boats on sea, or vehicle on roads (car, human), as shown in Fig. 1.

 Fig. 1: Examples of car and bike detection from UAV image sequences

 The analysis of images obtained from small aerial platform poses many challenges due to rapid platform motion, image instability and the relatively small size of the object of interest signatures within the resulting imagery. Recent approaches for the automatic detection of vehicles based on using multiple trained cascaded Haar classifiers is proposed [Gaszczak2011]. [Grabner2008] proposed a method to detect car from aerial images. The main contribution of [Grabner2008] is a new online boosting algorithm for efficient car detection. However, car detecting in [Grabner2008] is still objects. In this works, we will expend [Stephan2014] and [Grabner2008] for a robust object detection in an UAV image sequences.

[Stephan2014] proposed a framework includes still object segmentation (such as road, building, sky). We utilize such results in a framework of Bayesian net such as car object usually located in road, boat usually located in sea regions. To learn such object (car, boat), an online learning scheme (such as online boosting [Grabner2008]) is adopted.

 

Work description:

Theory:

  • Study UAV acquisitions schemes of the imaginary
  • Study image features extraction and description using color, textures, and SIFT
  • Study online boosting algorithms
  • Study classification utilizing Bayesian net algorithms

Practice:

  • A method to take advantages of Bayesian net and online boosting
  • A method embeds the results to applications on real environments such as deploying the proposed algorithms on the captured devices to detect and zoom-in an interesting object/ or counting number of the interested object

 Requirements:

 The students who have a fairly good knowledge about image processing and C++ programming are privileged. 

  

Student profile

This subject is dedicated to Vietnamese students as well as foreigner students at Master degree of Signal and Image processing option. 

   
Supervisors/contacts

Vu Hai, Computer Vision Department, MICA. Email: hai.vu at mica.edu.vn

Le Thi Lan, Computer Vision Department, MICA. Email: thi-lan.le at mica.edu.vn

 

References:

[1]   [Stéphane2014] Stéphane Lathuilière, Hai Vu, Thi-Lan Le, Thanh-Hai Tran Hung Dinh Tan, "Semantic Regions Recognition in UAV Images Sequence", in the Proceeding of the Sixth International Conference on Knowledge and System Engineering (KSE 2014), Hanoi, Vietnam, Oct., 2014

[2]   [Gąszczak2014], Anna Gaszczak et al , “ Real-time People and Vehicle Detection from UAV Imagery”, in the Proceeding of the 2011 Intelligent Robot and Computer Vision XXVIII, pp., 1-13.

[3]   [Grabner2008] Helmut Grabner, Thuy Thi Nguyen, Barbara Gruber, Horst Bischof, “On-line boosting-based car detection from aerial images”, ISPRS Journal of Photogrammetry and Remote Sensing, 2008