A computationally efficient recursive processing for multi-frame track-before-detect

Dynamic programming based track before detect (DP-TBD) is a batch processing method, which jointly processes number K of consecutive frames of measurements to improve the detection performance of dim targets. Due to its batch processing manner, sliding window (SW) processing is usually required to track long continuous time target trajectories, e.g., radar tracking. Every time when a new measurement is received, SW moves forward by one step to include the latest K frames of measurements and then update the existing target tracks and detect new targets in the light of extra information from new data. This means that K frames have to be processed every measurement time, and each frame needs to be repeatedly processed K times in K consecutive SWs, leading to a heavy computational complexity. In this paper, we present a computationally efficient approximation to the original SW processing method. Similar to a recursive filter, e.g., particle filter, it only uses the latest frame to update the target state, and each frame only needs to be processed once. Our analysis suggests that the proposed approximation admits negligible performance degradation but achieves significant reduction of computational cost, especially when K is large. Finally various numerical examples are used to demonstrate the robustness and efficiency of the proposed method.