Dynamic Programming-Based Multiple Point Target Detection Using K-means Clustering Algorithm

The problem of detecting multiple point targets that provide a level of situation awareness for unmanned aerial vehicles is addressed. The proposed tracking system, based on the track-before-detect approach, is designed to track and detect multiple targets from a sequence of a vision sensor under low SNR conditions. The system achieves multiple point target detection in three steps. The first step is morphological filtering process based on grayscale morphology for extracting intensive point-like features within image frame. Such filters are derived from combinations of dilation and erosion operations. The second step is target detection and tracking based on a dynamic programming approach. The dynamic programming approach accumulates scores of the pixels from the image sequence of morphological filter outputs along possible target trajectories. The scores for the potential target trajectories can be accumulated by considering the temporally and spatially uncorrelated noise and smoothly moving targets with only gradually changes in direction and speed. The decision of the target presence and position is made in the third step with threshold parameters set to achieve appropriate probabilities of detection and false alarm. In this step, K-means algorithm is used for identifying position and number of targets in two-dimensional space. The proposed track-before-detect approach using K-means clustering algorithm is applied to several image sequences containing different scenarios and noise conditions.