Mahfuzul Haque

Improved Gaussian mixtures for robust object detection by adaptive multi-background generation

Adaptive Gaussian mixtures are widely used to model the dynamic background for real-time object detection. Recently the convergence speed of this approach is improved and a relatively robust statistical framework is proposed by Lee (PAMI, 2005). However, object quality still remains unacceptable due to poor Gaussian mixture quality, susceptibility to background/foreground data proportion, and inability to handle intrinsic background motion. This paper proposes an effective technique to eliminate these drawbacks by modifying the new model induction logic and using intensity difference thresholding to detect objects from one or more believe-to-be backgrounds. Experimental results on two benchmark datasets confirm that the object quality of the proposed technique is superior to that of Leepsilas technique at any model learning rate.

On Stable Dynamic Background Generation Technique Using Gaussian Mixture Models for Robust Object Detection

Gaussian mixture models (GMM) is used to represent the dynamic background in a surveillance video to detect the moving objects automatically. All the existing GMM based techniques inherently use the proportion by which a pixel is going to observe the background in any operating environment. In this paper we first show that such a proportion not only varies widely across different scenarios but also forbids using very fast learning rate. We then propose a dynamic background generation technique in conjunction with basic background subtraction which detected moving objects with improved stability and superior detection quality on a wide range of operating environments in two sets of benchmark surveillance sequences.

Perception-Inspired Background Subtraction

Developing universal and context-invariant methods is one of the hardest challenges in computer vision. Background subtraction (BS), an essential precursor in most machine vision applications used for foreground detection, is no exception. Due to overreliance on statistical observations, most BS techniques show unpredictable behavior in dynamic unconstrained scenarios in which the characteristics of the operating environment are either unknown or change drastically. To achieve superior foreground detection quality across unconstrained scenarios, we propose a new technique, called perception-inspired background subtraction (PBS), which avoids overreliance on statistical observations by making key modeling decisions based on the characteristics of human visual perception. PBS exploits the human perception-inspired confidence interval to associate an observed intensity value with another intensity value during both model learning and background-foreground classification. The concept of perception-inspired confidence interval is also used for identifying redundant samples, thus ensuring the optimal number of samples in the background model. Furthermore, PBS dynamically varies the model adaptation speed (learning rate) at pixel level based on observed scene dynamics to ensure faster adaptation of changed background regions, as well as longer retention of stationary foregrounds. Extensive experimental evaluations on a wide range of benchmark datasets validate the efficacy of PBS compared to the state of the art for unconstraint video analytics.

A hybrid object detection technique from dynamic background using Gaussian mixture models

Adaptive background modelling based object detection techniques are widely used in machine vision applications for handling the challenges of real-world multimodal background. But they are constrained to specific environment due to relying on environment specific parameters, and their performances also fluctuate across different operating speeds. On the other side, basic background subtraction (BBS) is not suitable for real applications due to manual background initialization requirement and its inability to handle repetitive multimodal background. However, it shows better stability across different operating speeds and can better eliminate noise, shadow, and trailing effect than adaptive techniques as no model adaptability or environment related parameters are involved. In this paper, we propose a hybrid object detection technique for incorporating the strengths of both approaches. In our technique, Gaussian mixture models (GMM) is used for maintaining an adaptive background model and both probabilistic and basic subtraction decisions are utilized for calculating inexpensive neighbourhood statistics for guiding the final object detection decision. Experimental results with two benchmark datasets and comparative analysis with recent adaptive object detection technique show the strength of the proposed technique in eliminating noise, shadow, and trailing effect while maintaining better stability across variable operating speeds.

Panic-driven event detection from surveillance video stream without track and motion features

Modern surveillance systems are becoming highly automated in terms of scene understanding and event detection capabilities, and most existing methods rely on track-and motion-based features for event classification and anomaly detection. However, trajectory-based methods fail in public scenarios due to frequently loosing the object tracks, while the capabilities of motion-based methods are limited in detection of direction and velocity related anomalies. In this paper, a novel feature extraction and event detection method is presented without using any track and motion features where event discriminating characteristics are discovered from the dynamics of multiple temporal features extracted from foreground blobs and then confined in support vector machine based models for real-time event detection. Experimental results on benchmark datasets show that the proposed method can successfully discriminate panic-driven events like sudden split, runaway, and fighting from usual events.

Rotational assisted endoscopic retrograde cholangiopancreatography in patients with reconstructive gastrointestinal surgical anatomy

AIM To evaluate the success rates of performing therapy utilizing a rotational assisted enteroscopy device in endoscopic retrograde cholangiopancreatography (ERCP) in surgically altered anatomy patients. METHODS Between June 1, 2009 and November 8, 2012, we performed 42 ERCPs with the use of rotational enteroscopy for patients with altered anatomy (39 with gastric bypass Roux-en-Y, 2 with Billroth II gastrectomy, and 1 with hepaticojejunostomy associated with liver transplant). The indications for ERCP were: choledocholithiasis: 13 of 42 (30.9%), biliary obstruction suggested on imaging: 20 of 42 (47.6%), suspected sphincter of Oddi dysfunction: 4 of 42 (9.5%), abnormal liver enzymes: 1 of 42 (2.4%), ascending cholangitis: 2 of 42 (4.8%), and bile leak: 2 of 42 (4.8%). All procedures were completed with the Olympus SIF-Q180 enteroscope and the Endo-Ease Discovery SB overtube produced by Spirus Medical. RESULTS Successful visualization of the major ampulla was accomplished in 32 of 42 procedures (76.2%). Cannulation of the bile duct was successful in 26 of 32 procedures reaching the major ampulla (81.3%). Successful therapeutic intervention was completed in 24 of 26 procedures in which the bile duct was cannulated (92.3%). The overall intention to treat success rate was 64.3%. In terms of cannulation success, the intention to treat success rate was 61.5%. Ten out of forty two patients (23.8%) required admission to the hospital after procedure for abdominal pain and nausea, and 3 of those 10 patients (7.1%) had a diagnosis of post-ERCP pancreatitis. The average hospital stay was 3 d. CONCLUSION It is reasonable to consider an attempt at rotational assisted ERCP prior to a surgical intervention to alleviate biliary complications in patients with altered surgical anatomy.

Robust Background Subtraction Based on Perceptual Mixture-of-Gaussians with Dynamic Adaptation Speed

In this paper, we propose a new background subtraction technique based on perceptual mixture-of-Gaussians (PMOG). Unlike numerous variants of the classical MOG based approach [1], which can ensure reliable detection result only in known operating environments through proper parameter tuning, PMOG shows superior detection performance across dynamic unconstrained scenarios without any tuning. This is due to PMOGs intrinsic capability of exploiting several perceptual characteristics of human visual system for better understanding of the operating environment to avoid blind reliance on statistical observations. Furthermore, the proposed technique dynamically varies the model adaptation speed, i.e., learning rate, based on observed scene statistics for faster adaptation of changed background and better persistency of detected foreground entities. Comprehensive experimental evaluation on a number of standard datasets validates the robustness of the technique compared to the state-of-the-art.

Background Subtraction for Real-Time Video Analytics Based on Multi-hypothesis Mixture-of-Gaussians

Robust background subtraction (BS) is essential for high quality foreground detection in most video analytics systems. Recent BS techniques achieve superior detection quality mostly by exploiting the complementary strengths of multiple background models or processing stages. Consequently, these techniques fail to meet the operational requirements of real-time video analytics due to high computational overhead where BS is just the primary processing task. In this paper, we propose a new BS technique, named multi-hypothesis mixture-of-Gaussians (MH-MOG), suitable for real-time video analytics. The essential idea is to maintain a single background model based on perception-aware mixture-of-Gaussians and then, generating multiple detection hypotheses with different processing bases. Finally, only during the detection stage, the complementary strengths of the hypotheses are exploited to achieve superior detection quality without significant computational overhead. Comprehensive experimental evaluation validates the efficacy of MH-MOG.

Abnormal Event Detection in Unseen Scenarios

Event detection in unseen scenarios is a challenging problem due to high variability of scene type, viewing direction, nature of scene entities, and environmental conditions. Existing event detection approaches mostly rely on context-specific tuning and training. Consequently, these techniques fail to achieve high scalability in a large surveillance network with hundreds of video feeds where scenario specific tuning/training is impossible. In this paper, we present a generic event detection approach where the extracted low-level features represent the global characteristics of the target scene instead of any context-specific information. From the temporal evolution of these context-invariant features over a timeframe, a fixed number of temporal features are extracted based on the periodicity of significant transition points and associated temporal orders. Finally, top-ranked temporal features are used to train binary classifier-based event models. In this approach, supervised training and exhaustive feature extraction are required only once while building the target event models. During real-time operation in unseen scenarios, event detection is performed based on the trained event models by extracting the required features only. The proposed event detection approach has been demonstrated for abnormal event detection in completely unseen public place scenarios from benchmark datasets without additional training and tuning. Furthermore, the proposed event detection approach has also outperformed recent optical flow based event detection technique.