With the increasing popularity of deep learning on edge devices, compressing large neural networks to meet hardware requirements resource-constrained devices became a significant research direction. Numerous compression methodologies are currently being used reduce memory sizes and energy consumption networks. Knowledge distillation (KD) is among such it functions by using data samples transfer knowledge captured model (teacher) smaller one (student). However, due various reasons, original...

Local structures of shadow boundaries as well complex interactions image regions remain largely unexploited by previous detection approaches. In this paper, we present a novel learning-based framework for region recovery from single image. We exploit local edges using structured CNN learning framework. show that label information in classification can improve consistency over pixel labels and avoid spurious labelling. further propose formulate shadow/bright measure to model among regions....

Data-Free Knowledge Distillation (KD) allows knowledge transfer from a trained neural network (teacher) to more compact one (student) in the absence of original training data. Existing works use validation set monitor accuracy student over real data and report highest performance throughout entire process. However, may not be available at distillation time either, making it infeasible record snapshot that achieved peak accuracy. Therefore, practical data-free KD method should robust ideally...

Local structures of shadow boundaries as well complex interactions image regions remain largely unexploited by previous detection approaches. In this paper, we present a novel learning-based framework for region recovery from single image. We exploit the local edges using structured CNN learning framework. show that label information in classification can improve consistency results and avoid spurious labelling. further propose formulate shadow/bright measure to model among regions. The...

Model-free object tracking is still challenging because of the limited prior knowledge and unexpected variation target object. In this paper, we propose a feature learning algorithm for model-free multiple tracking. First, pre-learn generic features invariant to diverse motion transformations from auxiliary video data by using deep network anto-encoder. Then, adapt pre-learned according objects respectively in multi-task manner. We treat adaptation each as one single task. simultaneously...

Texture and color are two primitive forms of features that can be used to describe a scene. While conventional local binary pattern (LBP) texture based background subtraction performs well on rich regions, it fails detect uniform foreground objects in large background. As such, information complement feature. In this study, we propose incorporate feature Improved Hue, Luminance, Saturation (IHLS) space introduce an adaptive scheme automatically adjusts the weight between similarities pixel's...

With the proliferation of advanced computer vision techniques, machine vision-based Intelligent Transportation System (ITS) will be next wave technologies for better traffic management and monitoring. In this paper, we present a reconfigurable collaborative multi-camera framework vehicle detection license plate recognition. The synchronization between multiple cameras enable us envision in providing real-time recognition best usage system resources while bolstering high rate rate. addition,...

This paper presents a new keypoint-based approach to near-duplicate images detection. It consists of three steps. Firstly, the keypoints are extracted and then matched. Secondly, matched voted for estimation affine transform based on an invariant ratio normalized lengths. Finally, further confirm matching, color histograms areas formed by in two compared. method has advantage handling case when there only few keypoints. The proposed algorithm been tested Columbia dataset conducted...

This paper proposes a novel method for tracking persons based on the principal colors of human objects. First, an efficient object representation method, color (PCR), is proposed. Asymmetric similarity measures are then proposed representation. These asymmetric could be used to evaluate matching between two individuals as well visual evident individual in group. An algorithm or groups described. The has been tested using image sequences containing multiple moving frequently gathering and...

Shape and motion are two most distinct cues observed from human actions. Traditionally, K-Nearest Neighbor (K-NN) classifier is used to compute crisp votes multiple separately. The then combined using linear weighting scheme. Usually, the weights determined in a brute-force or trial-and-error manner. In this study, we propose new classification framework based on sum-rule fusion of fuzzy K NN classifiers. Fuzzy K-NN capable producing soft votes, also known as membership values. Based Bayes...

In this paper we present a method to detect vehicles on the road for intelligent transport application. proposed method, statistical property of background pixels are defined, where foreground objects current frame detected by subtracting w.r.t. color model and edge model. Color considered here as these two features provide complementary advantages. On top pixel-based step pixel models, region is incorporated, segmentation also subject its appearance in spatial domain. Clustering performed...

Traditional image stitching algorithms use transforms such as homography to combine different views of a scene. They usually work well when the scene is planar or camera only rotated, keeping its position static. This severely limits their in real world scenarios where an unmanned aerial vehicle (UAV) potentially hovers around and flies enclosed area while rotating capture video sequence. We utilize known geometry along with recorded trajectory create cylindrical images captured given...

Background subtraction is an essential technique for automatic video analysis. The main idea to construct and update the model of background scene. Foreground pixels are detected if they deviate from a certain extent. can consist color, texture gradient information [1]. In this paper, we focus on both color information. proposed feature based local binary pattern (LBP), while represented by (LCP). LBP known work well rich regions invariant subtle illumination variations, but it inefficient...

Merging and splitting of objects cause challenges for visual tracking. This is due to observation ambiguity, object lost, tracking errors when are close together. In this paper, we propose a method combine the joint probabilistic data association (JPDA) particle filter maintain tracks objects. The results JPDA employed improve model in filter. Based on ability handling missing detections clutter JPDA, can be maintained after merging or splitting. Conversely, also improves performance by...

Unattended object detection is a crucial task in visual surveillance systems. However, it challenging handling false alarms and miss rate. In this paper, two-stage method for the unattended proposed where first stage tries to detect all possible objects prevent detections by considering attributes of such as staticness, foregroundness, abandonment. This called proposal stage. second stage, our reduces with candidates obtaining from using deep learning similarity matching between background...

Visual surveillance systems use more and cameras in order to cover wider areas reduce blind spots. Cameras placement configuration usually depends on the area be monitored size of objects scene. Video analytics also require a minimal get detailed features or people. Most vision-based focus detection tracking people However, it is often meaningful describe with high-level information such as hair style, carrying bag other attributes. In perform this close view required. paper, collaborative...

This paper describes the design and implementation of autonomous real-time motion recognition on a Personal Digital Assistant. All previous such applications have been non required user interaction. The motivation to use PDA is test viability performing complex video processing an embedded platform. application was constructed using representation technique for identifying patterns Hu Moments. approach based upon temporal templates (Motion Energy History Images) their matching in time. done...

In many intelligent surveillance systems there is a requirement to search for people of interest through archived semantic labels. Other than searching typical appearance attributes such as clothing color and body height, information whether person carries bag or not valuable provide more relevant targeted search. We propose two novel fast algorithms sling backpack detection based on the geometrical properties bags. The advantage proposed that it does require shape from human silhouettes...