This paper proposes an algorithm for motion detection with the assumption of a still camera. Sigma-delta filtering is applied to measure the motion likelihood of each pixel. Then, the variation rate and temporal activity are calculated and compared to find the moving pixels. Pixels with higher temporal activity are labeled as moving pixels and saved in the motion images. For detection accuracy, only the edge pixels appearing in both the source and motion images are kept. Hybrid reconstruction is then applied to those pixels to reconstruct the whole moving object using forgetting morphological operators and morphological erosion. The performance of the algorithm is further enhanced by relevance feedback, which only applies sigma-delta filtering to the moving pixels in the source images. To deal with complex scenes, the algorithm computes multiple confidence values for multiple regions of an image. This improves the accuracy and adaptability of the background estimation. The algorithm is tested on five video surveillance sequences (three traffic, one indoor with two people, and one outdoor with many people). Experimental results show that the algorithm can efficiently separate the foreground moving objects and the background.

The proposed algorithm has low computational cost and a low memory requirement. It is suitable for parallel processing. The algorithm could be applied in video surveillance applications, where the detection of moving objects is of great interest and importance. This paper is recommended for researchers or engineers in the area of motion detection, object segmentation, and pattern recognition.