Kernelized correlation tracker on smartphones
Section snippets
Tracker benchmark
Ways of measuring the tracking performance are explained in [3], which describes three ways of measuring the tracking accuracy. One method is the OPE (One-Pass Evaluation), which is used in our work to compare tracking performances. This type of precision plot measures the average precision for a video sequence from the beginning to the end. With it, the percentages of frames of a sequence are measured, which do not exceed certain error thresholds. The location error threshold is the Euclidean
Color features
In this paper, we took a closer look into two sorts of color features. With both variants of color features, the image is first divided in square-sized cells and for each cell a histogram is calculated, similar to the HOG features.
Implementation
The KC tracking algorithm is implemented for the Android OS. The KC algorithm and the features (HOG, color) are implemented as a shared library in C/C++ with the NDK for Android. That way, APIs like OpenGL ES and OpenCL and other libraries, that e.g. SoC manufacturers offer, can be used very easily.
Performance comparison of color features to HOG features with MATLAB
Both color feature variants were implemented for MATLAB as built-in C functions, but without the use of vector registers (e.g. SSE) for even shorter runtimes. As the color features were implemented for MATLAB, one can also use the MATLAB implementation of the KC tracker with these features. This makes it also possible to compare the tracking performance using these color features with the HOG features.
The tracking performance (see Section 1.1) with the first variant of color features was
Conclusion
In this paper we showed that the KC tracker, which reaches a very good precision, can be implemented and run on smartphones with at least 20–30 FPS, if the available hardware (CPU, GPU, …) is used appropriately. Furthermore, we presented two approaches to color features, where the second variant reaches a very good performance with the used tracker, which is at least as good or even better as with HOG features. The combination of HOG and color features does significantly improve the precision
Future work
GPUs in smartphones are getting more powerful with every generation. Therefore, a fast GPU implementation for FFT transforms with any sample sizes could accelerate the tracking frame rates significantly. With the new Android camera API for image processing in Android 5.0, maybe higher camera frame rates than 30 FPS could be used in Android applications. Currently, high camera frame rates are only available for video captures (up to 120 FPS) but not within apps. To improve the detection of a
Acknowledgment
The authors would like to thank Prof. Dr. Franz Kurfess for his helpful comments/proofreading.
References (15)
- et al.
Exploiting the circulant structure of tracking-by-detection with kernels
- et al.
High-speed tracking with kernelized correlation filters
IEEE Trans. Pattern Anal. Mach. Intell.
(2015) - et al.
Online object tracking: A benchmark
- et al.
Struck: Structured output tracking with Kernels
- et al.
Histograms of oriented gradients for human detection
- et al.
Object detection with discriminatively trained part based models
IEEE Trans. Pattern Anal. Mach. Intell.
(2010) From rigid templates to grammars: Object detection with structured models
(2012)