Zacharias Kandylakis, Konstantinos Karantzalos, Anastasios Doulamis, Nikos Doulamis
Abstract: Although, cutting-edge frame hyperspectral sensors can currently acquire hypercubes at video rates with low spatial resolution, they offering enhanced discrimination capabilities for the characterization of subtle spectral features and important object reflectance properties. To this end, a generic framework was designed, developed and validated for multiple object tracking in hyperspectral video sequences. The background estimation was efficiently addressed through advanced scale space filtering and dimensionality reduction. The detection of the moving objects was performed on the reduced representation for low computational complexity. The object recognition task was based on certain spectral and geometric features which were associated with a rule-based classification. The experimental results appear promising and indicate the efficiency of the developed approach. |
Multiple Object Tracking |
|
| A flowchart of the developed algorithm for the automated multiple object tracking in hyperspectral video sequences |
 |
| The spectral bands of the raw hypercube require radiometric and geometric corrections. The checkerboard visualization demonstrate the effectiveness of the registration procedure. |
 |
| The raw, anisotropically smoothed (AML), the principal component of the lower dimensional representation (PCA) along with the estimated background are shown for (i) Top: frame #510 at 644nm and (ii) Bottom: frame #800 at 865nm. |
 |
| Experimental results from the ParkingFills hyperspectral video dataset. After the application of the developed method- ology the detected multiple objects have been classified and labelled correctly. |