14-05-2012, 04:11 PM
Various tracking methods
[attachment=22016]
Abstract
Various tracking methods have been developed to track objects with different degrees or levels of tracking ability. The ability or performance of each tracking method is dependent on the feature or data that is being used for tracking purpose. The ability of a tracking method can be measured by utilizing tracking metrics to give an indication of the tracking ability of an algorithm. This paper offers some insights into the issues and similarities of performance measurement reporting of video tracking algorithms and proposes a method in assessing the robustness of a video tracking algorithm. The proposed metric introduces another measure to measure the consistency of a tracking algorithm. The work presented in this paper shows that using only one metric to measure the tracking performance is inadequate. The proposed metric presented in this ¬paper shows that the utilization of multiple metrics such as tracking success rate and tracking consistency or robustness would give a better indication of the tracking ability of a tracking algorithm used in video surveillance.
Introduction
Current implementations of tracking and surveillance systems have provided valuable information and assistance in monitoring large areas. These systems typically utilize human operators to determine suspicious human behavior and to manually track people of interest over an extended area using an array of cameras. The limitation or difficulty for a human operator to constantly monitor cameras or to track people over an array of cameras had brought about the merging of technologies of cameras and computers. This in turn has lead to the growth of intelligence in the surveillance systems being developed. Along with these intelligent surveillance systems, tracking algorithms have been developed. Each tracking algorithm that has been developed is based on improving the performance of previous algorithms by utilizing different parameters or features [1]. Metrics are measures of the ability of an algorithm in processing signals ranging from still images, audio signals [2] to video signals. In the area of video surveillance, which covers the action of tracking, a target tracking algorithm would have to be developed. To determine the performance of a tracking algorithm, a performance metric needs to be able to determine its correctness or accuracy. In addition to determining correctness or accuracy, other metrics can also be used to characterize the performance such as tracking consistency (TC).
There have been various evaluation metrics that were developed to evaluate the performance of a tracking algorithm[3],[4],[5],[6],[7],[8]. Most of the published sets of agreed metrics utilized ground truth, which will be explored in the next section. There are alternatives to using ground truth to evaluate the performance of a tracking algorithm and a number of these have been identified by Ellis [9]. The alternatives include a method that requires each target to carry a mobile global positioning receiver to record 4D trajectory information that could be correlated directly with results of the video dataset. Another method cited takes advantage of the capability of automatically identifying targets in overlapping multi-camera views [10],[11],[12],[13],[14] and a third method uses synthetic image sequences to assess the performance of the algorithm [15].
In general, the objective of each tracking method was to ensure the ability of tracking or maintaining an accurate track of selected objects or targets of interest. The performance of each algorithm is based on the features used for tracking the selected targets. Tracking metrics are measures that indicate the performance or the ability of the tracking algorithms to successfully track the selected targets. In looking at the metric developed by different groups, there are similarities in terms of the measurement of particular performance metrics. Also, not all metrics developed are able to completely determine the performance of a tracking algorithm and as such other supporting metrics that measure certain other characteristics are required, in order to give better reporting of the tracking ability of an algorithm.
.BACKGROUND
Ground Truth
The term ground truth is typically used in the field of aerial photography, satellite imaging, and remote sensing. The primary purpose of these activities is to gather data from a distance and the term ground truth refers to the data that are being collected. In a simpler form, ground truth refers to "what is actually on the ground that needs to be correlated with the corresponding features in the scene (usually as depicted in a photo or image)" [16]. A typical example is crop classification using high-resolution satellite images by Omkar et al. [17]. Another example is the classification and error assessment of Landsat TM imagery [18].
Requirements for Effective Performance Analysis
Ellis's approach to performance evaluation takes into consideration how the algorithms deal with different physical conditions in the scene, for example, unrelated motion, weather conditions, and lighting. The performances of algorithms are assessed using ground truth. The proposed approach compare the tracked data to the marked data, in order to determine whether data are target position, 2D shape model, or classification of some description.
Automatic Evaluation System on Object Surveillance
Schlogl et al. [15] proposed an evaluation framework called Automatic Evaluation System on Object Surveillance (AESOS). AESOS comprises a set of error metrics and video reference data. The proposed framework demonstrated how motion segmentation and tracking can be used to determine the operational range of and entire surveillance system. As this paper is concerned with tracking in video surveillance, this section will focus only on the points salient to tracking evaluation. The proposed AESOS system generates scenes by selecting the number of objects, their trajectories, and velocities. Simple sequences are used for the evaluation of the operational range.
[attachment=22016]
Abstract
Various tracking methods have been developed to track objects with different degrees or levels of tracking ability. The ability or performance of each tracking method is dependent on the feature or data that is being used for tracking purpose. The ability of a tracking method can be measured by utilizing tracking metrics to give an indication of the tracking ability of an algorithm. This paper offers some insights into the issues and similarities of performance measurement reporting of video tracking algorithms and proposes a method in assessing the robustness of a video tracking algorithm. The proposed metric introduces another measure to measure the consistency of a tracking algorithm. The work presented in this paper shows that using only one metric to measure the tracking performance is inadequate. The proposed metric presented in this ¬paper shows that the utilization of multiple metrics such as tracking success rate and tracking consistency or robustness would give a better indication of the tracking ability of a tracking algorithm used in video surveillance.
Introduction
Current implementations of tracking and surveillance systems have provided valuable information and assistance in monitoring large areas. These systems typically utilize human operators to determine suspicious human behavior and to manually track people of interest over an extended area using an array of cameras. The limitation or difficulty for a human operator to constantly monitor cameras or to track people over an array of cameras had brought about the merging of technologies of cameras and computers. This in turn has lead to the growth of intelligence in the surveillance systems being developed. Along with these intelligent surveillance systems, tracking algorithms have been developed. Each tracking algorithm that has been developed is based on improving the performance of previous algorithms by utilizing different parameters or features [1]. Metrics are measures of the ability of an algorithm in processing signals ranging from still images, audio signals [2] to video signals. In the area of video surveillance, which covers the action of tracking, a target tracking algorithm would have to be developed. To determine the performance of a tracking algorithm, a performance metric needs to be able to determine its correctness or accuracy. In addition to determining correctness or accuracy, other metrics can also be used to characterize the performance such as tracking consistency (TC).
There have been various evaluation metrics that were developed to evaluate the performance of a tracking algorithm[3],[4],[5],[6],[7],[8]. Most of the published sets of agreed metrics utilized ground truth, which will be explored in the next section. There are alternatives to using ground truth to evaluate the performance of a tracking algorithm and a number of these have been identified by Ellis [9]. The alternatives include a method that requires each target to carry a mobile global positioning receiver to record 4D trajectory information that could be correlated directly with results of the video dataset. Another method cited takes advantage of the capability of automatically identifying targets in overlapping multi-camera views [10],[11],[12],[13],[14] and a third method uses synthetic image sequences to assess the performance of the algorithm [15].
In general, the objective of each tracking method was to ensure the ability of tracking or maintaining an accurate track of selected objects or targets of interest. The performance of each algorithm is based on the features used for tracking the selected targets. Tracking metrics are measures that indicate the performance or the ability of the tracking algorithms to successfully track the selected targets. In looking at the metric developed by different groups, there are similarities in terms of the measurement of particular performance metrics. Also, not all metrics developed are able to completely determine the performance of a tracking algorithm and as such other supporting metrics that measure certain other characteristics are required, in order to give better reporting of the tracking ability of an algorithm.
.BACKGROUND
Ground Truth
The term ground truth is typically used in the field of aerial photography, satellite imaging, and remote sensing. The primary purpose of these activities is to gather data from a distance and the term ground truth refers to the data that are being collected. In a simpler form, ground truth refers to "what is actually on the ground that needs to be correlated with the corresponding features in the scene (usually as depicted in a photo or image)" [16]. A typical example is crop classification using high-resolution satellite images by Omkar et al. [17]. Another example is the classification and error assessment of Landsat TM imagery [18].
Requirements for Effective Performance Analysis
Ellis's approach to performance evaluation takes into consideration how the algorithms deal with different physical conditions in the scene, for example, unrelated motion, weather conditions, and lighting. The performances of algorithms are assessed using ground truth. The proposed approach compare the tracked data to the marked data, in order to determine whether data are target position, 2D shape model, or classification of some description.
Automatic Evaluation System on Object Surveillance
Schlogl et al. [15] proposed an evaluation framework called Automatic Evaluation System on Object Surveillance (AESOS). AESOS comprises a set of error metrics and video reference data. The proposed framework demonstrated how motion segmentation and tracking can be used to determine the operational range of and entire surveillance system. As this paper is concerned with tracking in video surveillance, this section will focus only on the points salient to tracking evaluation. The proposed AESOS system generates scenes by selecting the number of objects, their trajectories, and velocities. Simple sequences are used for the evaluation of the operational range.