Illumination Invariant Mean-Shift Tracking Gargi Phadke, Rajbabu Velmurgan* Indian Institute of technology Bombay Mumbai,India,400076 gargiphadke@ee.iitb.ac.in rajbabu@ee.iitb.ac.in * Abstract Visual tracking is a critical task in surveillance and ac- tivity analysis. One of the major issues in visual target tracking is variations in illumination. In this paper, we propose a novel algorithm based on discrete cosine trans- form (DCT) to handle illumination variations, since illumi- nation variations are mainly reflected in the low-frequency band. For instance, low illumination in a frame leads to low value DC coefficient as vias versa. We modify DC co- efficient to achieve illumination invariance. Average of DC coefficients of particular numbers of neighboring frames of current frame is taken. The correction in DC coefficient is performed using maximum eigen value of the image co- variance matrix over N frames. The videos with corrected illumination are than used to track objects of interest using the Mean shift algorithm. The proposed algorithm is tested on an exhaustive database. The results demonstrate signifi- cantly improved tracking. 1. Introduction Robust and real-time tracking is a challenging problem in computer vision for applications in surveillance, moni- toring and activity analysis. It is cumbersome to track a moving object with changes in illumination. In [5] pro- poses a robust and low complexity mean shift tracker, it fails when an object undergoes changes in illumination con- ditions. Some of the existing techniques can deal with mi- nor changes in illumination but fail with sudden changes in illumination for example [10]. In [2] multi resolution tech- niques is described for tracking the targets in changes in illumination, but it fails when there are drastic changes in illumination. All these methods are pixel based enhance- ment. When illumination is low or high in video, the en- ergy oriented method in spatiotemporal form is explained in [3], but cannot handle sudden changes in illumination. [17] described multi-space method for tracking targets un- der changes in illumination. All these methods are for use- ful only when illumination is not dark. In[10] background weighted histogram is used for target model, but illumina- tion changed drastically this method fails. In[4] author has proposed a method to enhance image using gaussian mix- ture model but it has drawback that it cannot enhance to- tally low illuminated image because the enhancement de- pends upon histogram of image. Wavelet fusion is used to enhance the image [8][11], but it also depends upon maxi- mum value of intensity. 2. Proposed Method Stable illumination condition is an important prerequi- site for successful tracking. We propose to modify the D.C. coefficient of DCT of input frames with low illumination in order to create a video with unvarying luminance. First, we transform every video frame from RGB to YUV plane where Y-plane represents the luminance parameter. Loga- rithmic transformation is applied on this Y-plane to separate luminance from reflectance followed by DCT . The first coefficient of DCT is the D.C. coefficient, which is propor- tional to the illumination level in the frame, i.e. higher the luminance, higher will be the D.C. value. Taking K consec- utive frames into consideration, we modify the D.C. values to get uniform illumination throughout, mainly by improv- ing luminance of darker frames. This is achieved with the help of the proposed correction factor which is computed using statistical properties of every frame. Mean shift is then used for tracking the targets, in enhanced illumination invariant video. 3. Logarithmic and Discrete Transform Illumination of surveillance video is not predictable. It depends upon the real life factors like the time and the place where the video is captured. Methods for illumination- invariant image processing are given in [12]. The basic concept for illumination invariance follows from the im- age formation model in which image intensity of a pixel at a location [x, y] in an image f [x, y], is assumed to be the 407 978-1-4673-5052-5/12/$31.00 ©2012 IEEE 407 978-1-4673-5054-9/13/$31.00 ©2013 IEEE