Efficient event-detection for smart cameras using motion cues

Abstract

The number of surveillance cameras installed in most environments is increasing rapidly, indicating the importance of visual surveillance in the modern society. This provides an opportunity to use intelligent vision-based algorithms to enhance the functionality of these cameras. Such vision-based algorithms are targeted towards event detection in the scene, i.e., the system should be capable of detecting an event of interest in the scene. However, event detection systems have practical constraints in accuracy and implementation. The event detection system should be accurate while requiring little delay, area and power. In this dissertation, we propose efficient event detection algorithms based on motion cues. For sparsely populated environments, where any motion in the scene is an event of interest, background subtraction can be used. For crowded environments, crowd anomaly detection schemes can be employed. In Chapter 1, the motivation and methodology of this research is described. In Chapter 2, we present a new background subtraction scheme named EBSCam. We first show that the background model in sequential estimation schemes fluctuates rapidly with noisy input. We develop analytical models for the case of GMM, which show the effect of background model fluctuation on foreground classification error. Afterwards we propose EBSCam, which has a relatively stable background model. Also, we propose EBSCam with Poisson Mixture Model (EBSCam-PMM). Both EBSCam and EBSCam-PMM outperform standard background subtraction schemes in background subtraction performance. The memory bandwidth, delay, area and energy consumption of EBSCam and EBSCam-PMM are superior compared to state-of-the-art implementations of background subtraction. These properties make EBSCam and EBSCam-PMM suitable candidates for surveillance applications. In Chapter 3, we propose a new crowd anomaly detection algorithm based on outlier rejection. We assume that pixels in a local neighborhood with similar direction of motion belong to the same object. Pixels which do not adhere to the dominant direction of motion can be rejected and should not contribute to the feature of the pixel. We also propose a modification to the standard K-means algorithm, which significantly improves crowd anomaly detection performance. The proposed method outperforms numerous state-of-the-art methods. Also, we propose a hardware implementation of the proposed algorithm, which is much faster, requires low energy and area compared to reported implementations. Concluding remarks over the dissertation are presented in Chapter 5.