Investigates Image data fusion techniques that combine image and track data from multiple sensors to achieve improved accuracies and more specific inferences than could be achieved by the use of a single sensor alone. Our aim is to explore the state-of-the-art image processing algorithms for achieving effective data fusion as in:
Flight Target Tracking
Contents
1. Introduction
2. Main Algorithm and Principle
3. Demo
1. Introduction
Flight target tracking using Electro-optical sensor is a challenging issue in surveillance system due to adverse environments such as cloudy weather. Furthermore, Flight target has fast speed and generally has small image information. The conventional centroid tracking methods are suffered from many clutters according to incorrect segmentation. Otherwise, the conventional meanshift tracking methods are suffered from occlusion problem, etc. To solve these problems of each method, we propose the switching conditions between centroid tracking and meanshift tracking for utilizing advantages of each method.
2. Main Algorithm and Principle
- Overall Flow
1. Set to default method to centroid tracking
2. If the centroid tracking phase, perform the switching condition test for centroid tracking result at each frame.
3. If fail the test, switch to meanshift tracking method.
4. If the meanshift tracking phase, perform the switching condition test for meanshift tracking result at each frame.
5. If the image was recovered background environment suitable for centroid tracking, switch to centroid tracking.
6. If the all test is fail, use only Kalman filter prediction result.
- Switching Conditions
A. Centroid-to-Meanshift
Condition 1 : No observations. Observations do not exist in the validation region
Condition 2 : Too many observation. Many observations are observed in the validation region
- Due to the segmentation failure or occlusion.
Condition 3 : Local standard deviation test
- The characteristic of flight target is darker value in the small region.
- Local standard deviation value of target region is measured higher value than value of background.
B. Meanshift-to-Centroid
Condition 1 : Similarity measure between estimated region and reference region
- Use the histogram intersection measure.
Condition 2 : Inverse of the Centroid-to-meanshift test
- The case that recover the environment suitable for centroid tracking.
3. Demo
For the performance evaluation, we use the data generated by simulator.
