Accepted 1.0 Savantic letter 1(6) False alarm in outdoor environments
Accepted 1.0 Savantic letter 2(6) Table of contents Revision history 3 References 3 1 Introduction 4 2 Pre-processing 4 3 Detection, tracking, classification 5 4 Conclusions 6
Accepted 1.0 Savantic letter 3(6) Revision history Status Version Date Person Comment Accepted 1.0 13-03-12 Claes Orsholm Reviewed 0.2 13-03-11 Christofer Aourell Comments added Not reviewed 0.1 13-03-05 Sergei Prasalovich First draft References [1] Shadow detection and removal, Savantic letter (2012) 2.2. [2] Tracking, why and how?, Savantic letter (2012) 2.1. [3] K. Garg and S.K. Nayar - Detection and Removal of Rain from Videos, Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. [4] A.K. Tripathi and S. Mukhopadhyay - Removal of rain from videos: a review, Signal, Image and Video Processing, 2012. [5] P.T. Barnum et al. - Analysis of Rain and Snow in Frequency Space, International Journal of Computer Vision (2010) 86, pp. 256274. [6] K. Liu et al. - A Joint Optical Flow and Principal Component Analysis Approach for Motion Detection, IEEE International Conference on Acoustics Speech and Signal Processing (2010), pp. 1178-1181. [7] Y. Benabbas et al. - Motion Pattern Extraction and Event Detection for Automatic Visual Surveillance, EURASIP Journal on Image and Video Processing (2011) 7. [8] X. Zhao and G. Medioni - Robust Unsupervised Motion Pattern Inference from Video and Applications, Proceedings of IEEE International Conference on Computer Vision (2011), pp. 715-722. [9] V. Saligrama et al. - Video Anomaly Identification - A statistical approach, IEEE Signal Processing Magazine (2010) 27 (5), pp. 18-23.
Accepted 1.0 Savantic letter 4(6) Abstract This letter studies triggers of false alarms in outdoor video surveillance systems and overviews potential solutions based on the latest techniques in image analysis. The following two approaches are discussed: a) filtering by pre-processing of video streams and b) filtering by detection, tracking and classification. False alarms due to weather and small animals and birds is the main focus of this letter. 1 Introduction False alarm is one of the most important issues in any automated surveillance system. Video surveillance systems are the most common type. It consists of a Closed-Circuit Television (CCTV) platform where the signal from one or several video cameras is transmitted through a server to a limited set of monitors. In order to perform automated surveillance controls such systems also require a motion detection platform for the video stream analysis and pre- or post-processing. The efficiency of such surveillance systems are defined by the detection probability, which itself is dependent on the detection sensitivity and False Alarm Rate (FAR). Decreasing FAR while keeping sensitivity as high as possible is the ultimate goal of a common surveillance task. Indoor environments have typically controlled lighting conditions and static space arrangements. Therefore it is often possible to achieve reliable performance by limiting the cameras Region-Of-Interest (ROI) and/or by tuning detection sensitivity. Outdoor environments, on the contrary, are more dynamic and often result in varying scenes and changing lighting conditions. Also taking weather into account as an extra complication factor, performing even simple surveillance tasks become much more complex. In such environments a more intelligent filtering is required in order to achieve reliable detection. The following are the most common sources of false alarm in outdoor environments: Weather (rain, snow, wind, clouds) Animals (dogs, cats, rats etc) and birds Lighting conditions (reflections, shadows) The goal of this letter is to analyse the above sources of false alarms in outdoor video surveillance systems and to overview possible solutions. This letter focuses mostly on studies of false alarm due to weather conditions such as rain and snow and due to small animals and birds. Studies of false alarm from varying lighting conditions (shadows) is presented elsewhere [1]. 2 Pre-processing Rain and snow are the most typical examples of bad weather that vitally effect the FAR in outdoor surveillance. Such weather conditions can vary widely in their physical properties and thus in the resulting visual effects they can produce in images. Rain and
Accepted 1.0 Savantic letter 5(6) snow are classified as dynamic types of weather conditions in contrast to static types like fog, mist and haze. The constituent particles of dynamic weather conditions are typically larger compared to the static ones. One might even be able to distinguish single particles in images. While falling with high velocities such particles (rain drops) can produce so called steaks resulting in its turn in sharp intensity changes in images due to reflection, refraction and scattering of light. One of the proposed solutions to the problem of false alarm due to weather conditions like rain or snow is based on pre-processing video streams and filtering weather effects out before detecting for policy violation. In some approaches, photometrical properties of environment and dynamics of rain are used, for example, to build models of how rain drops effect captured images. A raindrop is viewed there as an optical lens that refracts and reflects light thus producing a motion blur effect or a rain steak. Based on these and some additional assumptions an algorithm for removal of rain steaks has been developed by [3]. While achieving some satisfactory results for filtering effect of rain steaks from light rain, this approach however fails in a number of other more realistic conditions like steady or heavy rain when the constraints of photometric model are not fulfilled anymore. Among other methods proposed are: Kalman filter-based method, shape characteristics-based method, low-latency removal method, histogram model-based, probabilistic model-based. A good review including performance comparison of all the above methods is described in [4]. An interesting idea has also been considered in [5], namely to go from the temporal domain to the frequency domain in order to detect and filter blurring effects of rain in each frame taking into account global properties of the entire scene. This method works even for snow and in dynamic scenes with moving background and camera. A drawback of this approach lies however in unpleasant artefacts caused in image space as a result of pre-processing. 3 Detection, tracking, classification Small animals and birds are another very common trigger of false alarms in outdoor environments. These types of targets are almost impossible to filter away from scenes by pre-processing of video streams as discussed above in the case of rain. The solution to this problem therefore lies in filtering during detection and/or tracking. It can be divided into two steps: 1) detection, where video stream is analysed frame-by-frame and intelligent motion detection is performed, including first stage filtering of false alarms using a series of tests based, for example, on the target location, area or shape information; 2) tracking, where detected targets are followed and there motion is tested against another set of rules allowing second stage filtering to be performed based on such characteristics of movement like trajectory and speed, for example. Object detection and tracking involving blobs - pixel representations of physical objects obtained by background subtraction, is a widely used method in surveillance applications. The blobs are usually tracked using Kalman or particle filters. This method however performs poorly in varying lighting conditions and when the tracked objects are
Accepted 1.0 Savantic letter 6(6) occluded. Alternative approaches use feature points or points of interest like corners, edges or other features for tracking. A technique known as Optical Flow (OF) can be applied then to track detected features. This method proved to be very effective in applications where physical objects are hard to single out like, for example, in extremely crowded scenes where the size of moving objects on images is just a few pixels. An interesting approach is presented in [6] where OF is combined with Principal Component Analysis (PCA) for motion detection resulting in improved performance with reduced FAR for videos with either static or dynamic background. OF methods are in general more computationally effective, however when it comes to more sophisticated tasks they also fail at performing in real time. Tracking performance can be improved by using motion patterns instead of trajectories as shown in [7, 8]. In [8], for example, blobs and OF approaches are combined together by extracting a set of tracklets from each detected motion blob which is then used to learn motion patterns. Tracking of individual objects is facilitated by using known motion pattern information thus improving tracking and overall detection performance. More information on how good detection and tracking effects false alarm can be found in [2]. Upon successful feature detection and tracking, an object s trajectory can be classified using trained data sets. Anomalous behaviour detection in surveillance applications is still a very challenging task and choice of techniques is very much dependent on specific scenarios. A state-of-the-art review of video anomaly detection techniques is described very well in [9]. A new approach is introduced there based on statistical activity analysis. Its main idea is to locate relevant activities prior to higher-level analysis such as feature extraction, tracking and classification. This allows avoiding the clutter of entire scenes by focusing on located abnormalities i.e. narrowing ROI of analysed scenes. 4 Conclusions False alarm in video surveillance of outdoor environments has been discussed. False alarm from rain and snow can be minimized by filtering images during pre-processing stage. Other types of false alarms, for example, events triggered by small animals or birds has to be taken care of by improving filtering during actual detection and tracking of potential targets. A number of solutions have been discussed based on motion tracking and classification as well as on alternative statistical methods where relevant activities are identified prior to feature extraction and tracking. False alarm in outdoor video surveillance however remains to be a very complex problem to solve. There is no best or all-in-one solution to it. The choice of techniques depends very strongly on the details of the scenery and requirements of the particular surveillance system.