In recent years, there has been an increase in demand for intelligent automatic surveillance systems to detect abnormal activities at various places, such as schools, hospitals, prisons, psychiatric centers, and public gatherings. The availability of video surveillance cameras in such places enables techniques for automatically identifying violent actions and alerting the authorities to minimize loss. Deep learning-based models, such as Convolutional Neural Networks (CNNs), have shown better performance in detecting violent activities by utilizing the spatiotemporal features of video frames. In this work, we propose a violence detection model based on 3D CNN, which employs a DenseNet architecture for enhanced spatiotemporal feature capture. First, the video’s redundant frames are discarded by identifying the key frames in the video. We exploit the Multi-Scale Structural Similarity Index Measure (MS-SSIM) technique to identify the key frames of the video, which contain significant information about the video. Key frame identification helps to reduce the complexity of the model. Next, the identified video key frames with the lowest MS-SSIM are forwarded to 3D CNN to extract spatiotemporal features. Furthermore, we exploit the Convolutional Block Attention Module (CBAM) to increase the representational capabilities of the 3D CNN. The results on different benchmark datasets show that the proposed violence detection method performs better than most of the existing methods. The source code for the proposed method is publicly available at https://github.com/venkateshakula19/violence-detection-using-keyframe-extraction-and-CNN-with-attention-CBAM

In recent years, the growing use of biometric recognition systems in various applications has increased the need to protect the biometric templates recorded in multiple databases. Due to their consistency and uniqueness, iris recognition systems have significantly outperformed other biometrics. Directly stored Iris templates on a central server constitute a privacy and security risk. To address this, we will generate a cancelable template that can be stored instead of the original. In the event of a security breach, we will discard the stored template and generate a new iris template. This research employs the Weber Local Descriptor (WLD) technique to create a multi-instance iris biometric system. Left and right iris images are initially acquired and normalized using the USIT toolkit. We generate a feature vector from the normalized image using WLD. The obtained feature vector is then normalized using L1 normalization. The vector of normalized features is then projected onto a median filter to generate a cancelable template. Experiments are conducted on the IIT Delhi iris database, and the results are optimistic compared to previously published research.

Human Activity Recognition (HAR) based on sensor networks is of paramount importance in the fields of body area networks, ubiquitous and pervasive computing. HAR is widely used in applications such as health monitoring, medical care, smart homes etc. With the advent of sensor networks and the fast-growing waveform data in the technologically developing modern world, the traditional feature engineering methods are becoming more obsolete. Deep Learning methods are very beneficial as they are efficient in feature extraction, helps in modelling the sensor data systematically and improving the performance of complex human activity recognition. Taking advantage of deep learning techniques, we propose a model based on Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). To this we integrate a special feature recalibration framework based on attention mechanism to perform human activity recognition. The model uses Inception Neural Network architecture with various kernel-based convolution layers to extract spatial features and Gated Recurrent Units (GRU) to model temporal / time-series features. Space and Channel based Squeeze and Excitation blocks (SCbSE) framework is used to recalibrate features to complete classification tasks of complex human activities. The proposed model is experimentally verified on two publicly available benchmark HAR datasets namely: Smartphone Dataset and Opportunity dataset. The performance of the model is analysed while comparing to that of the state-of-the-arts.