Infrared Physics & Technology 62 (2014) 59–64
Contents lists available at ScienceDirect
Infrared Physics & Technology journal homepage: www.elsevier.com/locate/infrared
Regular article
Infrared dim target detection technology based on background estimate Liu Lei ⇑, Huang Zhijian College of Electronic Engineering and Photoelectric Technology, Nanjing University of Science and Technology, Nanjing, China
h i g h l i g h t s Basic principles and the implementing flow charts of infrared dim target detection algorithms. Dim target detection experiments for IR images. Subjective and objective evaluation method for infrared dim target detection algorithms.
a r t i c l e
i n f o
Article history: Received 26 September 2013 Available online 9 November 2013 Keywords: Target detection Infrared small target Background estimate
a b s t r a c t Accurate and fast detection of infrared (IR) dim target has very important meaning for infrared precise guidance, early warning, video surveillance, etc. In this paper, two new algorithms – background estimate and frame difference fusion method, and building background with neighborhood mean method are presented. The basic principles and the implementing procedure of these algorithms for target detection are described. Using these algorithms, the experiments on some real-life IR images are performed. The whole algorithm implementing processes and results are analyzed, and those algorithms for detection targets are evaluated from the two aspects of subjective view and objective view. The results prove that the proposed method has satisfying detection effectiveness and robustness. Meanwhile, it has high detection efficiency and can be used for real-time detection. Ó 2013 Elsevier B.V. All rights reserved.
1. Introduction The main thought on infrared detection and tracking technology is as follows: Extracting target from pre-given infrared frame for operators, meanwhile, tracking object and predicting next position according to actual movement orbit. Presently, the operating principle of infrared object detection system is that infrared sensor receives infrared radiation from both object and background, and then generates real-time infrared image with different temperature distribution in terms of physics characteristics of target and background. As there are all kinds of disruptive factors triggered by system or environments, we have to revise real-time images before the image pro-processing to improve signal to noise ratio so as to realize target detection by using the correspondent detection algorithms one by one. It needs to make a correlation with current detected target and previous detection result as there may be falsetarget existing in the detection process. We can build up the movement orbit of correct result for the follow-up processing. Finally,
⇑ Corresponding author. Tel./fax: +86 25 84314969. E-mail address:
[email protected] (L. Lei). 1350-4495/$ - see front matter Ó 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.infrared.2013.10.010
we can realize detection and tracking with effective algorithm [1–5]. In this paper, the infrared target detection algorithms are discussed, namely, background estimate and frame difference fusion method, and building background method with neighborhood mean. Then we do some experiments by using this software platform. The implementing processes and results are analyzed. The experimental results show that these methods are efficient for the detection of the infrared targets. It has the great significance practical for the application. 2. The detection algorithm principle 2.1. Background estimate and frame difference fusion method First of all, one frame is considered as the background frame in background estimate method, after that, a new background is attained by background update mechanism, at the moment, we can get target information by subtraction between current frame and corresponding background frame. Although this method is simple and easy for hardware implementation and it is able to extract the characteristic information of targets. However, the background
60
L. Lei, H. Zhijian / Infrared Physics & Technology 62 (2014) 59–64
estimate method is sensitive to the natural conditions, once the intensity of natural light changes quickly or there are cloud shelters, it will have great influence on the testing results. This is a deadly deficiency which is hard to overcome for the algorithm, at the same time, this method is very dependent on efficient background model and its update mechanism. Secondly, the traditional frame difference method mentioned has its own shortage, it is easy to produce ‘‘cavity area’’ or ‘‘double image’’, so we would rather combined with the advantages of the two algorithms in the subject, so that it can get better experimental effect. After the experiment, it is proved that this algorithm can adapt to a variety of environment, and to some extent, it is superior to traditional two frame difference method and background estimate method [6–7]. The process of background estimate and frame difference fusion method is introduced as following: Firstly, setting a frame as the initialized background image Bi(x, y), then the current frame Ii subtracts the background image Bi, so we will get a difference image DBi(x, y), then we compare each gray value in the difference image with the threshold T, which can be expressed by the mathematical formula (1):
d ¼ jIi ðx; yÞ Bi ðx; yÞj
ð1Þ
Ii(x, y) means the gray levels of pixels in current frame, Bi(x, y) means the gray levels of pixels in background frame, d means difference image. Thresholding on formula (1):
DBi ðx; yÞ ¼
1; d P T 0; d < T
ð2Þ
DBi(x, y) means the pixel gray scale value after binaryzation, T is judging threshold. Now combining the binary images obtained by background difference method and the binary image obtained by improved three frames difference method, using mathematical expression to show as follows:
( C i ðx; yÞ ¼
1; DBi ðx; yÞ [ Bi ðx; yÞ ¼ 1 ð3Þ 0; DBi ðx; yÞ [ Bi ðx; yÞ–1
Bi(x, y) means the three frames difference result, DBi(x, y) means the background difference result, Ci(x, y) means the infrared target finally extracted from the image. The whole operation process can be shown in flow chart Fig. 1. From Fig. 1, we can see that in the process by using the background estimated method and frame difference method to detect target, first of all, the original video is decomposed into frames sequence, and changed it into gray image; and then the first frame gray image is considered as the initial background frame, and we do subtraction operation between the current frame gray image and the corresponding background frame, then threshold on the difference results and do logical ‘‘and’’ operation for the binary image of background difference method and the binary image of improved three frames difference method, if the result is ‘‘1’’, it is set as a foreground pixel, otherwise it is set as a background pixel, after that we need to update the background image.
Fig. 1. The flow chart of background estimate and frame difference fusion method.
61
L. Lei, H. Zhijian / Infrared Physics & Technology 62 (2014) 59–64
2.2. Building background method with neighborhood mean value The three target detection algorithms mentioned above are developed from traditional target detection algorithm. Now we propose a new detection algorithm, namely the method of building background with neighborhood mean. The effect of this method is better for infrared video than for color video. The main principle is we firstly build background image with the pixel’s neighborhood mean value, and then, current frame subtracts corresponding background frame to get difference results, at the same time, we update background image through efficient background update mechanism constantly to adapt the surrounding environment, so that the follow-up of detection and tracking results are more accurate, then we do binaryzation operation on the difference result. Because of the existence of noise, so we need to filter the binary image by using mathematical morphology operation, and then get the designated target, the procedure is as follows: (1) Background pixel initialization: first of all, we should get the first frame of continuous image sequence, then calculate the sum of gray scale values of N21 (N = 2n + 1) pixels surrounding a pixel Ii(x, y) and take the average value Bi(x, y) by the formula (4), finally use the average value to replace background gray scale value of the pixel [7].
Bi ðx;yÞ ¼
1 2
N 1
X Ii ðk;tÞðx n 6 k 6 x þ n; y n 6 t 6 y þ nÞ k;t
ð4Þ (2) Calculating the difference value Ri(x, y) between the pixel Fi(x, y) in current frame and the pixel Bi(x, y) in corresponding background by the formula (5),
Ri ðx; yÞ ¼ jF i ðx; yÞ Bi ðx; yÞj
ð5Þ
(3) Setting the proper threshold according to the dynamic threshold segmentation method. (4) Binaryzation: thresholding on the result obtained from the third step by the formula (6),
Di ðx; yÞ ¼
1; Ri ðx; yÞ > T 0; Ri ðx; yÞ 6 T
ð6Þ
(5) Updating the background pixel, combining the background pixel in former frame and the background pixel in current frame with weight factor a by the formula (7),
Biþ1 ðx; yÞ ¼ ð1 aÞ Bi ðx; yÞ þ a Biþ1 ðx; yÞ
ð7Þ
(6) Morphological filtering processing on the binaryzation result to extract the target. The specific process of this algorithm can be shown in Fig. 2. First of all, the original video is decomposed into frames sequence, and changed into gray image; And then we use the mean value of
Fig. 2. The flow chart of building background method with the neighborhood means value.
62
L. Lei, H. Zhijian / Infrared Physics & Technology 62 (2014) 59–64
Fig. 3. Original infrared frames.
pixels around a pixel in the first frame to construct initial background frame, and do difference operation with current frame and background frame, then compare the difference operation result with the threshold, if it is larger than threshold, it is thought as the foreground pixel, otherwise it is the background pixel; In order to improve signal-to-noise ratio of the picture, we take morphological filtering operation for the final target detection. 3. Results analysis We make use of target detection and tracking software to test an infrared .avi video file and a color .avi video file, and evaluate the results for the algorithms from the subjective and objective aspects. Infrared video is of .avi format. The whole image frames number is 200, video sample size is 24 bits, the frame size is 160 120 pixels, and frame rate is 25 frames per second [8]. Fig. 3 shows a few frames from the original infrared video which is used for experimental process. The color video is of .avi format. The whole image frames number is 489, video sample size is 24 bits, the frame size is 276 206 pixels, and frame rate also is 25 frames per second. The original color video frame is shown in Fig. 4. 3.1. The results analysis for background estimation and difference frame fusion method Figs. 5 and 6 show that the detection results for color and infrared video by using the background estimation and frame difference fusion method. We can see that for the infrared video the background estimation and frame difference fusion method can detect
the target clearly, but it has image blur, and a small amount of noise in the image. However, for color video, this method has better effect. 3.2. The results analysis for building background method with neighborhood mean Figs. 7 and 8 have presented the results of building background method with neighborhood mean for infrared and color video. We can see from Fig. 7 that building background method with neighborhood mean can detect infrared video target successfully, the three frames image displays the outline of two pedestrians, however, there a small amount of noise in the three frames image, which is decided by the size of pixels selected around. In Fig. 8, we do not detect the target in the 36th frame. The 168th frame also shows a very indistinct outline. Although we can detect pedestrians in the 248th frame, the targets are surrounded by a lot of noise, so building background method is better for infrared video than for color video after comprehensive comparison, and the algorithm can be set as a special research direction in the future for infrared dim-small target detection. 3.3. The performance evaluation for infrared target detection algorithms First of all, after we have observed the actual detection effect of these algorithms, we can conclude that the building background method with neighborhood mean value can be available for both infrared video and color video, while the building background method with neighborhood mean, only applies for infrared video.
Fig. 4. Original color video frame. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Fig. 5. The detection effect of background estimate and frame difference fusion method for infrared video.
L. Lei, H. Zhijian / Infrared Physics & Technology 62 (2014) 59–64
63
Fig. 6. The detection effect of background estimation and frame difference fusion method for color video.
Fig. 7. The detection effect of building background method with neighborhood mean for infrared video.
Fig. 8. The detection effect of building background method with neighborhood mean.
Table 1 The accuracy comparison of detection algorithms. Original video
Infrared video Color video
Detection algorithm Background estimation and frame difference fusion method
Building background method with neighborhood mean
90.6% 93.2%
83.9% 0%
Finally, we should calculate detection rate, which is the time spent on processing each frame. The smaller the detection rate is, the more frames image can be processed, and the higher processing efficiency is. The specific comparison data is shown in Table 2. From Table 2, it can be seen that background estimation and frame difference fusion method has the highest efficiency both for infrared video or color video, so it can be realized in real-time hardware process in the future research. 4. Conclusion
Table 2 The detecting rate of different algorithms (unit: ms/frame). Original video
Infrared video Color video
Detection algorithm Background estimation and frame difference fusion method
Build background method with neighborhood mean
46.54 49.24
69.10 –
Secondly, we should calculate target detection accuracy, which is the ratio of the real frames with targets detected and the total number of frames. The comparison results can be shown as Table 1. We can see from Table 1 that the background estimation and frame difference fusion method is more stable, it can get good effect for both infrared video and color video.
In this paper, some detection and tracking algorithms have been optimized and picked up in terms of implement possibility, advantages and shortcomings. Then we realize the detection and tracking of moving infrared object in infrared video and colored video by self-developed software. And an assessment on these algorithms is made. The testing results show the detection and tracking algorithms mentioned in the paper are effective to moving object, especially for infrared moving object. The results are meaningful for infrared detection and tracking application. Acknowledgements This work is sponsored by the National Natural Science Foundation of China (Grant No. 61101195), the Key Laboratory of Intelligent Perception and Systems for High-Dimensional Information (Nanjing University of Science and Technology), Ministry of Education (Grant No. 30920130122005), and the Scientific Research
64
L. Lei, H. Zhijian / Infrared Physics & Technology 62 (2014) 59–64
Foundation for the Returned Overseas Chinese Scholars, State Education Ministry (Grant No. 2013-693). References [1] M. Yasuno, S. Ryousuke, N. Yasuda, M. Aoki, Pedestrian, Detection and Tracking in Far Infrared Images, in: IEEE Conference on Intelligent Transportation Systems, 2005 (IEEE, 2005), pp. 182–187. [2] R.M. Liu, E.Q. Liu, J. Yang, Y. Zeng, F.L. Wang, Y. Cao, Automatically detect and track infrared small targets with kernel Fukunaga—Koontz transform and Kalman prediction, Appl. Opt. 46 (2007) 7780–7791. [3] K. Brunnstrom, B.N. Schenkman, B. Jacobson, Object detection in cluttered infrared images, Opt. Eng. 42 (2003) 388–399.
[4] A. Bal, M.S. Alam, ‘‘Dynamic target tracking with fringe adjusted joint transform correlation and template matching, Appl. Opt. 43 (2004) 4874–4881. [5] Guan, Q. Chen, G. Gu, W. Qian, Dim target detection based on optical flow histgram in low frame frequence in clouds background, Acta Opt. Sin. (28) (2008) 1496 (in Chinese). [6] Xin Wang, Zhenmin Tang, Modified particle filter-based infrared pedestrian tracking, Infrared Phys. Technol. 53 (4) (2010) 280–287. [7] Zhijian Huang, The Algorithm Study and Implementation on the Infrared Small Target Detecting and Tracking Technology Based on Open CV, Ph. M. Dissertation, Nanjing, 2012. [8] Zhang Junju, Tian Si, Chang Benkang, et al., Infrared Target Tracking in RealTime Video and its Implementation, in: International Symposium on Photoelectronic Detection and Imaging, 2007, Proc. of SPIE vol. 6621 66211S–1.