Motion Detection in UAV Videos by Cooperative Optical Flow and Parametric Analysis Masaharu Kobashi

1
Motion Detection in UAV Videos by Cooperative
Optical Flow and Parametric AnalysisMasahar
u Kobashi
2
Objectives

• Show contrast between simple research videos and
  UAV videos.
• Describe strengths and weaknesses of well-know
  conventional techniques.
• Propose a new system designed for robustness.

3
Overview of Presentation

1. Overview of motion analysis
2. UAV videos and their characteristics
3. Strengths and weaknesses of conventional
   techniques
4. Design of our new system
5. Performance of our system

4
What is motion analysis?

• Motion is a relative concept
• Relative to image plane (static camera)
• Relative to objects in the image
• Factors that make it difficult
• Changing extrinsic and intrinsic camera
  parameters (Pan, tilt, translation, rotation, and
  zooming)
• Illumination changes
• Overlap of multiple motions with
  semi-transparency (dust, fog, mist, glasses, etc.)

5
Classes of Motion Analysis Methods

• Detecting flow of regularly-shaped regions
• Optical flow, Patch/Block flow
• Detecting flow of irregularly-shaped regions
• Segmentation-based flow analysis
• Variable-window/block-based flow analysis
• Detecting difference between two frames in terms
  of an objective function
• Parametric motion analysis

6
Popular Videos for Motion Analysis
7
What are UAV videos?
8
Characteristics of UAV videos

• Noisier and of lower quality than standard videos
  
• Low bandwidth of UAV transmitter
• Quality of camera
• Constant changes of camera parameters (airplane
  motion, camera's pan, tilt, zooming)
• Often abrupt and quick changes
• Unrestricted natural scenes
• Including dimly lit scenes, deserts, fields

9
Contrast betweenResearch videos and UAV
videos(See following demo movies.)
10
Strengths and Weaknesses ofTechniques Related
to Motion Analysis

• Gradient-based Search
• Pyramid-Based Coarse-to-Fine Approach
• Optical-Flow-Based Motion Detection
• Parametric Motion Estimation
• Hybrid Approaches

11
Gradient-Based Search

• Hill-Climbing for global optimization based on
  local gradient information
• Examples
• Lucas-Kanade registration algorithm
• Light constancy equation
• Strength
• Fast (avoiding exhaustive search)
• Weakness
• Brittle (1-pixel support a pixel noise can
  derail it)

12
Pyramid-Based Coarse-to-Fine Approach

• Strengths
• Can summarize information compactly
• Remedy for range-limited gradient-based methods

13
Weaknesses of Pyramid

• Fixed coverage of child cells (Fixed
  partitioning)
• Same information to all child cells
• Vital information can be lost in summarizing
  process
• All parts of image required to use same number of
  levels

14
Results from Pyramid's Weaknesses

• The more levels, the more errors.
• Small motion areas suffer unnecessary errors due
  to the height of the pyramid needed for areas of
  greater motion.
• Motion of thin/small parts tends to be detected
  incorrectly.

15
Optical-Flow-Based Motion Analysis

• Steps
• Detection of flow for all pixels in each frame.
• Accurate segmentation of the flow into moving
  and stationary objects.

16
Strengths and Weaknesses of Optical-flow-Based
Motion Analysis

• Strength
• Dense motion information (for every pixel)
  available.
• Weaknesses
• Difficult to obtain accurate vectors for all
  pixels.
• Segmenting flow into multiple moving and static
  objects is difficult.

17
Difficulties in Finding Flow Discontinuity
18
Parametric Motion Estimation

• Steps
• Assume a model of transformation (Affine,
  Perspective).
• Compute transformation matrix of the static
  regions from the point correspondences of static
  regions.
• Find point correspondences between two frames f1
  and f2.
• Find difference between frame f1 and an image
  created by warping frame f2 by the
  transformation matrix.

19
Strengths and Weaknesses of Parametric Motion
Estimation

• Strengths
• Capable of precise matching between frames at
  all points even at sub-pixel precision.
• Unlike optical flow analysis, it can find
  different motion layers equally well in any
  transformation.
• Weaknesses
• Needs reasonably accurate point correspondences.
• Must know which points belong to static region.

20
Hybrid Approach

• Use of multiple motion detection techniques in a
  single system (e.g. optical flow detection
  parametric motion estimation)
• Examples
• Wang and Adelson's motion layers
• Our system

21
Overview of Our System

• Design philosophy
• Hybrid system based upon optical flow detector
  and parametric motion estimator
• Takes advantage of strengths of both and avoid
  their weaknesses
• Priority on robustness rather than efficiency
• System structure

22
Module 1 Optical Flow Detector

• Block matching with relatively large block
  (16x16)
• Use of Disk Balance Ratios (DBR)
• Use of Sum of Squared Error (SSD)

23
Disk Balance Ratios (DBR)

• Objective
• To describe the orientation of a region in terms
  of intensity.
• Definition

24
Limitations of SSD
25
Use of Sum of Squared Difference (SSD)

• Steps to find the flow vector of block i
• If the SSD distribution is too flat among
  candidates, make no decision. (Near homogeneous
  region.)
• If there are two or more candidates which have
  distinctively better SSD than the rest and flow
  vectors from block i to these candidates are far
  apart, make no decision. (Aperture problem.)
• Otherwise, find the candidate block having the
  least SSD. The vector from block i to that
  candidate is the flow vector of block i.

26
Module 2 Extraction of Reliable Flow

• Reliable flow is defined by the following
  reliability measure.

27
Module 3Finding Connected Reliable Flow Groups

• A variant of the single-link clustering
  algorithm.
• Connect adjacent reliable blocks if their flow
  vectors are similar in both angle and magnitude.

28
Module 4Finding Ground-Covering Flow Groups

• Steps
• If there is a group whose area is more than 40
  of the screen, the group is regarded to cover the
  ground only.
• If available, use the grouping information of the
  previous frame in the following way
• For a group g, if more than 70 of the area of
  g is covered by the ground region in its
  predecessor frame, then g belongs to the ground.
  Otherwise, make the decision by the next step.
• Select the set of groups that minimizes the SSD
  between the intensities of corresponding pixels
  between the template frame and the image created
  by warping the next frame by the inverse of the
  transformation matrix.

29
Computation of Transformation Matrix
30
(No Transcript)
31
Singular Value Decomposition (SVD)

• Merit
• Capable of handling even singular matrix

32
Module 5Evaluation of Difference Data

• Steps
• Compute the mean difference m of all pixels in
  the reliable ground blocks.
• Exclude the reliable ground blocks from the
  search region for the moving objects.
• In the remaining blocks find the pixels whose
  difference value is greater than m1.5s.
• Remove isolated small regions from the list of
  candidate regions.
• Mark all blocks which contains a candidate
  region as candidate blocks.
• Form connected components of candidate blocks.
  Each connected set of blocks is a moving object.
• For each connected set of blocks, output its
  bounding box to indicate the location of the
  moving object.

33
Output
34
Performance Comparison

• Lucas-Kanade plus pyramid v.s. Our system
• Michael Black's Algorithm v.s. Our system
• Simple research videos v.s. UAV videos

35
Lucas-KanadePyramid
36
Lucas-Kanade Pyramid
37
Black's Algorithm
38
Black's Algorithm