针对自然生境环境下高原鼠兔跟踪中，鼠兔毛色呈保护色与背景颜色相近以及运动随机的问题，构造了一种局部纹理差异性算子LTDC（local ternary difference count），来表征目标和背景之间的细微差异性，弥补了采用单一LTC（local ternary count）算子的不足。通过运动信息来判断鼠兔的运动模式，不同的模式采用不同的采样跟踪策略。把所构造的LTDC算子与R（red）G（green）B（blue）颜色信息相结合来表示目标，并把该目标表示模型引入到运动信息引导的高原鼠兔跟踪方法中。通过对采集的秋冬季节高原鼠兔视频图像进行测试，分析跟踪的成功率和误差，得到的 LTDC 纹理颜色模型的目标表示方法在鼠兔发生突变运动时，由于采用了运动信息引导的采样跟踪方式，能够有效地捕获突变目标，跟踪成功率达到97.93%。在鼠兔发生平滑运动时，尽管目标与背景颜色相近，依然能够稳定地跟踪目标，跟踪误差较小，误差波动范围也较小，误差均值为19.56，误差方差为74.24。试验结果表明：所提出的跟踪方法具有较强的目标与背景区分能力，在目标和背景颜色相近、运动复杂的场景中，能够较为准确地实现高原鼠兔目标的定位。%Plateau pika(Ochotona curzoniae) is one of the main biological disasters in the Qinghai-Tibet plateau and adjacent areas in China. Video-based animal behavior analysis is a critical and fascinating problem for both biologists and computer vision scientists. According to the color similarity betweenPlateau pika and the background, as well as the uncertainty and randomness of plateau pika motion in natural habitat environment, a new visual descriptor named the local texture difference operator LTDC was proposed to reflect the subtle differences between the plateau pika and the background. The LTDC operator, being more robust to target expression, made up for the deficiency of using single LTC. The LTDC operator discarded the structure information of local texture and retained the difference information of local texture, with low compute complexity. Aimed at the uncertainty and randomness of plateau pika motion, considering the prior knowledge that the position displacement between two adjacent frames was smaller in smooth movement and the position displacement between two adjacent frames was larger in abrupt motion, we extracted motion information between the adjacent frames using the frame difference method at first, then judged the movement mode of plateau pika by motion information, taking appropriate sampling tracking strategy to track plateau pika. If the mode was judged to be a smooth motion mode, we employed the Markov Chain Monte Carlo sampling tracking method based on the motion smoothness assumption. Else we adopted Wang-Landau Monte Carlo sampling tracking method used for abrupt motion tracking. Considering the fact of that object tracking method of motion-induction algorithm based on HSV color histogram usually had the deformation of inaccurate tracking or loss of target in the scenario where the color was similar between the background and the object, the LTDC operator was combined with RGB color information to characterize the object model, and the object model was embedded into the motion induction tracking framework for the plateau pika tracking. The test video collected the plateau pika activity behavior in the winter of 2014 in natural habitat environment, located in Qinghai-Tibet Plateau, eastern longitude 101°35′36″ -102°58′15″, northern latitude 33°58′21″ -34°48′48″. The video was totally 254 frames, with its size 320 pixels×240 pixels, and the fame rate 25 frames per second. The video feature was that the color was very similar between the plateau pika and background. Simultaneously, the plateau pika motion, being abrupt and occurring occasionally, was very stochastic. To test the tracking performance of the proposed method, we compared the tracking results obtained from proposed method with those of the motion-induction method and the WLMC method. Because the target representation was HSV color histogram in motion-induction method and the WLMC method, it was inclined to fail to track target. The target was lost at 39th frame with motion-induction method, and at 10th frame with WLMC method. But the tracking performance of proposed method using LTDC texture color model could locate the target all the time, even when the abrupt motion occurred. With the motion-induction method being compared with WLMC method, the tracking success rate of proposed method reached 97.93%, but the tracking success rate of motion-induction method and WLMC method were 31.82% and 24.79% respectively, which were 32.49% and 25.31% of the tracking success rate with proposed method. The error of the proposed method was smaller and the error fluctuation range was also smaller. The tracking stability of the proposed method was superior to that of motion-induction method and WLMC method. The error mean of proposed method was 62.79% and 67.24% of that calculated with the motion-induction method and WLMC method respectively. The error variance of proposed method was 19.74% and 19.66% of that obtained with the motion-induction method and WLMC method respectively, reducing by 80.26% and 80.34%. The experimental results show that the proposed tracking method has a strong distinguishing ability of target and background. The object can be accurately positioned under the scenario of color similarity between the object and the background, and the scenario of complex motion ways.
展开▼
