Today, machine learning and deep learning have paved the way for vital and critical applications such as abnormal detection. Despite the modernity of transfer learning, it has proved to be one of the crucial inventions in the field of deep learning because of its promising results. For the purpose of this study, transfer learning is utilized to extract human motion features from RGB video frames to improve detection accuracy. A convolutional neural network (CNN) based on Visual Geometry Group network 19 (VGGNet-19) pre-trained model is used to extract descriptive features. Next, the feature vector is passed into Binary Support Vector Machine classifier (BSVM) to construct a binary-SVM model. The performance of the proposed framework is evaluated by three parameters: accuracy, area under the curve, and equal error rate. Experiments performed on two different datasets comprising highly different context abnormalities accomplished an accuracy of 97.44% and an area under the curve (AUC) of 0.9795 for University of Minnesota (UMN) dataset and accomplished an accuracy of 86.69% and an AUC of 0.7987 for University of California, San Diego Pedistrain1 (UCSD-PED1) dataset. Moreover, the performance of the pre-trained network VGGNet-19 with handcrafted feature descriptors and with other CNN pre-trained networks, respectively, has been investigated in this study for abnormal behavior detection. The results demonstrated that VGGNet-19 has better performance than histogram of oriented gradients, background subtraction, and optical flow. In addition, the VGGNet-19 shows higher detection accuracy than other pre-trained networks: GoogleNet, ResNet50, AlexNet, and VGGNet-16.
Crowd counting considers one of the most significant and challenging issues in computer vision and deep learning communities, whose applications are being utilized for various tasks. While this issue is well studied, it remains an open challenge to manage perspective distortions and scale variations. How well these problems are resolved has a huge impact on predicting a high-quality crowd density map. In this study, a hybrid and modified deep neural network (U-ASD Net), based on U-Net and adaptive scenario discovery (ASD), is proposed to get precise and effective crowd counting. The U part is produced by replacing the nearest upsampling in the encoder of U-Net with max-unpooling. This modification provides a better crowd counting performance by capturing more spatial information. The max-unpooling layers upsample the feature maps based on the max locations held from the downsampling process. The ASD part is constructed with three light pathways, two of which have been learned to reflect various densities of the crowd and define the appropriate geometric configuration employing various sizes of the receptive field. The third pathway is an adaptation path, which implicitly discovers and models complex scenarios to recalibrate pathway-wise responses adaptively. ASD has no additional branches to avoid increasing the complexity. The designed model is end-to-end trainable. This integration provides an effective model to count crowds in both dense and sparse datasets. It also predicts an elevated quality density map with a high structural similarity index and a high peak signal-to-noise ratio. Several comprehensive experiments on four popular datasets for crowd counting have been carried out to demonstrate the proposed method’s promising performance compared to other state-of-the-art approaches. Furthermore, a new dataset with its manual annotations, called Haramain with three different scenes and different densities, is introduced and used for evaluating the U-ASD Net.
9 January 2024 in International Journal of Disaster Risk Reduction
Analyzing crowded environments has become an increasingly researched topic, largely due to its myriad practical applications, including enhanced video surveillance systems and the estimation of crowd density in specific settings. This paper presents a comprehensive method for progressing the study of crowd dynamics and behavioral analysis, specifically focusing on the classification of movement patterns. We introduce a specialized neural network-based classifier explicitly designed for the accurate categorization of various crowd scenes. This classifier fills a unique niche in the existing literature by offering robust and adaptive classification capabilities. To optimize the performance of our model, we conduct an in-depth analysis of loss functions commonly employed in multi-class classification tasks. Our study encompasses four widely-used loss functions Focal Loss, Huber Loss, Cross-Entropy Loss, and Multi-Margin Loss. Based on empirical findings, we introduce a Joint Loss function that combines the strengths of Cross-Entropy and Multi-Margin Loss, outperforming existing methods across key performance metrics such as accuracy, precision, recall, and F1-score. Furthermore, we address the critical challenge of class imbalance in motion patterns within crowd scenes. To this end, we perform a comprehensive comparative study of two leading oversampling techniques: the synthetic minority oversampling technique (SMOTE) and adaptive synthetic sampling (ADASYN). Our results indicate that ADASYN is superior at enhancing classification performance. This approach not only mitigates the issue of class imbalance but also provides robust empirical validation for our proposed method. Finally, we subject our model to a rigorous evaluation using the Collective Motion Database, facilitating a comprehensive comparison with existing state-of-the-art techniques. This evaluation confirms the effectiveness of our model and aligns it with established paradigms in the field.
27 January 2021 in Turkish Journal of Electrical Engineering and Computer Sciences
This paper proposes a hybrid steganographic ap-proach using the least significant bit (LSB) technique for grayscale images. The proposed approach uses both LSB match-ing (LSB-M) and LSB replacement to hide the secret data in images. Using hybrid LSB techniques increase the level of security. Thus, attackers cannot easily, if not impossible, extract the secret data. The proposed approach stores two bits in a pixel. The embedding rate can reach up to 1.6 bit per pixel. The proposed approach is evaluated and subjected to various kinds of image processing attacks. The performance of the proposed algorithm is compared with two other relevant techniques; pixel-value differencing (PVD) and Complexity Based LSB-M (CBL). Experimental results indicate that the proposed algorithm out-performs PVD in terms of imperceptibility. Also, it significantly outperforms CBL in two main features; higher embedding rate (ER), and more robust to most common image processing attacks such as median filtering, histogram equalization, and rotation.