1. DETECTION OF VEHICLES IN SHADOW AREAS USING COMBINED HYPERSPECTRAL AND LIDAR DATA M. Shimoni*; G. Tolt**; C. Perneel*** and J. Ahlberg** * Signal and Image Centre, Dept. of Electrical Engineering (SIC-RMA), Brussels, Belgium, ** FOI Swedish Defence Research Agency, Linköping, Sweden , *** Dept. of Mathematics, Royal Military Academy, Brussels, Belgium.

2. CASI-1500 382 nm to 1040 nm with 24 spectral bands, IFOV: 0.25 m.

3. Optech ALTM Gemeni LIDAR Spatial Res.: 0.25 m; Alt Res.: 0.10

5. STEP 1 : Shadow detection Training area SVM (force shadow/no shadow) Detect the interior of large area using distance transformation

6. Model of stationary European vehicle In the orthorectifed normalized Digital Surface Model (nDSM=DEM − DSM) and HIS data sets, a vehicle covers maximum 12 to 20 pixels in different delineations.

7. STEP 2 : Segmentation of vehicles in 3D LIDAR data Selected method: Mean Shift (MS) Major challenge: how to adaptively determine the kernel bandwidth in the spatial domain. For our purpose: a cylinder-shaped kernel is used, where the horizontal bandwidth h h (k) at each point varies and has to be determined based on joint height edge and pixel shape information (PSI): PSI is obtained through extending direction lines radiating from the central pixel based on spatial homogeneity PH d , PH d is the height difference between a centric pixel and every pixel on the direction lines. where L d (k) is the length of the d th direction line, T 1 is a predefined threshold for PH d (T 1 =0.75-1.50) T 2 is the maximal number of pixels allowed in the direction line (T 2 =20 pixels). Binary edge(x) indicates the presence of height edge.

8. Once all the direction lines are extended, the kernel bandwidth can be determined as follows: where with D being the total number of direction lines.

9. STEP 3 : Detection of spectral anomalies in shadows areas Selected method: Topological Anomaly Detection (TAD)* Processing steps: (1) The shadow areas were selected as ROI and been counted to assure the availability of at least 20X20 pixels. In negative case, this shadow area was not processed. (2) Normalisation : we normalize the shadow areas so that the brightest 10% of the pixels in the areas have Euclidean L2 norm equal to two and the darkest 1% have Euclidean L2 norm equal to one. (3) Background model: the shadow area pixels are chosen to model the background and the pairwise pixel distances between all pixels in the background set are computed. (4) Ranking each test pixel is compared to the pixels in the background set via the codensity metric,  k , which represents the radius of the smallest sphere enclosing k neighbors. * Basener, W., I. E. and Messinger, D., “Anomaly detection using topology,” Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery XIII SPIE 6565 (April 2007).

10. The TAD ranking of each pixel in the image is equal to the sum of the distances from the 3rd, 4th, and 5th nearest neighbors in the background pixels, i.e., The TAD rankings result in level sets of arbitrary topology, allowing detection of pixels in the holes of the convex hull of the background. Post processing: The TAD score is then thresholded in order to create a binary anomaly map. A vegetation index can be computed on the detected anomalies, so that the pixels containing biomass can be removed from the detections. The remaining anomalies are potential vehicles.

11. The TAD results after biomass removal

12. STEP 4 : Detection of occluded vehicles based on decision fusion Proposed method: shape/spectral integration (SSI) decision fusion algorithm Algorithm score assumption: The SSI algorithm exploits spectral target detection methods to nominate pixels as targets (vehicle), but uses shape information to boost target pixel scores above the scores of spectrally similar false alarms in the background by assuming that vehicles correspond to local height maxima.

13. Final results:

14. Conclusions: The SSI decision fusion algorithm was shown to outperform the MS and TAD anomaly algorithm mainly in deep shadow areas; The utility of shape information was shown to be a way to enhance spectral target detection in complex urban scene. The results illustrated that LIDAR is not independent of target-sensor geometry and that combined spectral information can be used to improve its detection in urban scene.

16. DETECTION OF VEHICLES IN SHADOW AREAS M. Shimoni*; G. Tolt**; C. Perneel*** and J. Ahlberg** * Signal and Image Centre, Dept. of Electrical Engineering (SIC-RMA), Brussels, Belgium, ** FOI Swedish Defence Research Agency, Linköping, Sweden , *** Dept. of Mathematics, Royal Military Academy, Brussels, Belgium.