A Texture-based Method for Detecting Moving Objects

Transcription

1 A Texture-based Method for Detecting Moving Objects Marko Heikkilä University of Oulu Machine Vision Group FINLAND

2 Introduction The moving object detection, also called as background subtraction, is one of the most critical components in many machine vision applications, especially in proactive ones. The goal is to extract the moving objects from the video frames captured from a camera. Moving Object Detection (Background Subtraction)

3 Applications of Moving Object Detection Automatic video surveillance Traffic monitoring People counting Military application Human-Machine interfaces Object tracking Etc.

4 Background Subtraction Roughly speaking, the background subtraction can bee seen as two-stage process as illustrated below. Foreground Detection Background Modelling

5 Background modeling The background modeling is the most important part of any background subtraction algorithm. The goal is to construct and maintain a statistical representation of the scene that the camera sees. Foreground Detection Background Modelling

6 Foreground Detection The comparison of the input frame with the current background model. The areas of the input frame that do not fit to the background model are considered as foreground. Foreground Detection Background Modelling

7 Requirements Background subtraction algorithm should be robust to Illumination changes. Multimodality of the background. Introduction of new stationary objects to the background. Removal of old stationary objects from the background. Shadows of the moving objects. Etc. Should operate in real-time.

8 The New Approach Instead of just using the pixel intensity, we chose to utilize image texture when modeling the background. We chose to use the Local Binary Patterns, LBP, as the texture measure because of its good properties: Invariant to monotonic changes in gray scale, which makes it robust against illumination changes. Computationally very fast. The algorithm operates under the assumption of a stationary camera with fixed focal length.

9 The New Approach The background is divided into blocks. We used partially overlapping blocks in our experiments. Each block is modeled as a group of K weighted LBP histograms: {x 1,x 2,,x K }. The background model is updated with the information of each new video frame, which makes the algorithm adaptive. The update procedure is identical for each block. x 1 x 2 x K

10 The New Approach 1. Calculate an LBP histogram x t for the block of the new video frame. 2. Compare the new block histogram x t against the existing K model histograms {x 1,x 2,,x K } by using the histogram intersection as the distance measure. 1. If none of the model histograms is close enough to the new histogram, the model histogram with the lowest weight is replaced with the new histogram and is given a low initial weight. 2. If a model histogram close enough to the new histogram was found, the bins of this histogram are updated as follows: x k,t [i] = α b x t [i] + (1 - α b )x k,t-1 [i] 0 α b 1 (1) Furthermore, the weights of the model histograms, {ω 1, ω 2,, ω K }, are updated as follows: ω k,t = (1 - α w ) ω k,t-1 + α w M k,t 0 α w 1 (2) where M k,t is 1 for the matched histogram and 0 for the others.

11 The New Approach 3. Next, we need to decide which of the histograms of the model are most likely produced by the background processes. We use the persistence of the histogram as an evidence for this. Because the persistence of the k th model histogram is directly related to its weight ω k,t, the model histograms are sorted in decreasing order according to their weights. As a result the most probable background histograms are on the top of the list. 4. As a last phase of the updating procedure, the first B model histograms are selected to be the background model as follows: ω 1,t + ω 2,t + + ω B,t > T B 0 T B 1 (3)

12 The New Approach Foreground detection is achieved via comparison of the new block histogram x t against the existing B background histograms {x 1,x 2,,x B } selected at the previous time instant. If a match is not found, the block is considered to belong to the foreground. Otherwise, the block is marked as background.

13 Experiments The performance of the proposed algorithm was evaluated using several test sequences. Both indoor and outdoor videos were included. Real-time performance. The performance was compared to four algorithms representing state-ofthe-art pixel and block based approaches.

14 Results The new approach proved to be tolerant to Illumination changes. Multimodality of the background. Introduction of new stationary objects to the background. Removal of old stationary objects from the background. Real-time performance: For the parameter values used in the tests, a frame rate of 28 fps was achieved. We used standard PC with a 1.8 GHz AMD Athlon XP CPU processor and 512 MB of memory in our experiments. The image resolution was 320 times 240 pixels.

15 Results

16 Results We manually labeled five frames from all of the test sequences and calculated the accumulated sums for the true positives and the true negatives reported by the algorithms for the test sequences. The overall performance of our approach seems to be better than the performance of the comparison methods. True Positives True Negatives This Work [2] [3] [4] [5]

17 Further Development Frame rate improvement by optimizing the implementation. Study possibilities to extend the algorithm to handle the moving camera case too.

18 Thanks!