Features of pretrained network models (9, 10) are depicted in Table 1 below.

We have done an extensive survey on how researchers worked on anomaly-related datasets like UCF-Crime and its subsets like HR-Crime, XD-Violence, UCSD Anomaly Detection Dataset (crowd anomaly), Shanghai-Tech, LAD (large anomaly dataset), Avenue, CAVIAR, PETS-2016, and so on, and we have listed their methods and achieved accuracy.

Sultani et al. (4) from the University of Central Florida made the UCF-Crime dataset. It has a total of 1,900 videos in 13 classes like Abuse, Arson, and so on and a total video length of 128 h.

Mascorro et al. (5) used 3D-CNN to detect abnormal behavior on shoplifting cases. They introduced a new concept to analyze precrime scenes:

Sultani et al. (4) used deep neural networks with multiple instance learning to classify real-world anomalies including accidents, explosions, conflicts, abuse, arson, and so on. The AUC for their product is 75.41%. They obtained accuracies of 23.0 and 28.4% using C3D and TCNN Architecture, respectively.

Sabokrou et al. (2) used CNNs with 3D deep autoencoders to detect irregularities in videos.

Ullah et al. (6) utilized an approach that used 15 consecutive frames of video to construct a feature vector, which was then fed into a multilayer bidirectional LSTM to differentiate anomalous occurrences. They were 85.53% correct. On UCF-Crime, the VGG-19 with multilayer BD-LSTM achieved an accuracy of 82%. On UCF-Crime, the concept V3 with multilayer BD-LSTM achieved 80% accuracy.

Hasan et al. (3) created a convolutional autoencoder (Conv-AE) framework for scene reconstruction and then estimated reconstruction costs for abnormality detection.

Dubey et al. (7) suggested the 3D deep Multiple Instance Learning with ResNet (MILR) approach as well as a novel proposed ranking loss function. With that new ranking loss function, they obtained an AUC of 76.67%.

In their suggested technique, Nasaruddin et al. (11) used a strong background subtraction to extract motion and identify the locations of attention areas. Eventually, a 3D CNN receives the output areas. They used C3D (convolution 3-dimensional) to their full advantage, developing a deep convolutional network to discern between typical and anomalous occurrences. Their locality learning model achieved an accuracy of 99.25%.

AUC results based on publicly available codes (8, 9) are shown in Table 2 below.

Results given in anomaly detection in video sequences: A benchmark and computational model by Boyang et al. (8, 9).

The total video length of the anomaly part is 10 h 03 min. The total video length of the normal part is 62 min. The total video length of the criminal action detection dataset is 11 h 05 min. The average length of a video is 1 min 42 s. Snapshots of instances of different categories of action from our dataset are shown in Figure 1 below.

We have made our own convolutional Long Short Term Memory Gated Recurrent Unit model (conv-LSTM-GRU) and Long Recurrent Convolutional Network (LRCN) model to predict.

Used python libraries:

A Time Series is an assortment of data congregated over time. In such instances, a model based on LSTM, a Recurrent Neural Network architecture, is an attractive solution. The previous concealed state is sent to the next phase in the sequence in this design. As a result, the network stores information based on past data and consumes it to make judgements. In other words, data order is crucial.

When working with photographs, a CNN architecture is the optimum option. Convolutional layers are used to extract essential elements from the picture. The output is joined to a fully coupled dense network after going through a series of convolutional layers. Conv-LSTM layers can be used in the situation of successive pictures. It is a recurrent layer like the LSTM, except that internal matrix multiplications are substituted with convolution operations. As a result, data passing through the Conv-LSTM cells keeps the original dimension.

GRUs and LSTM are quite similar. GRU uses gates to regulate the information flow, the same as LSTM. When compared to LSTM, they are quite new. They have a simpler design and provide certain improvements over LSTM because of this. In order to construct a new model to make predictions over a video as Time Series Data of a series of frames, we attempt to integrate the properties of Conv-LSTM and GRU. Figures 2, 3, respectively, depict Conv-LSTM (12) and GRU (13).

Long-term recurrent convolutional networks (LRCNs) are architectures that utilize CNNs for visual recognition and extend them to time-varying inputs and outputs. They examine visual inputs (potentially variable-length) and outputs into recurrent sequence models (LSTMs), resulting in variable-length predictions. The CNN and LSTM weights are shared, allowing scaling to any sequence length. The architecture of LRCN is depicted in Figure 4 (1).

The number of parameters in CONV-LSTM-GRU is given in Figure 5.

The design of the CONV-LSTM-GRU model is depicted in Figure 6 below.

The number of parameters used in LRCN is shown in Figure 7 below.

The design of LRCN model is presented in Figure 8 below.

From Figure 9, we can see clearly that the loss decreases as we increase the number of epochs; hence, we can conclude that the model has reached a global minimum solution. The validation loss also decreases along with the training loss; hence, we can say the model does not suffer from overfitting.

From Figure 10, we can see clearly that the accuracy increases as we increase the number of epochs; hence, we can conclude that the model is a near perfect fit and does not suffer from overfitting.

From Figure 11, we can see clearly that the loss decreases as we increase the number of epochs; hence, we can conclude that the model has reached a global minimum solution. The validation loss also decreases along with the training loss; hence, we can say the model does not suffer from overfitting.

From Figure 12, we can see clearly that the accuracy increases as we increase the number of epochs; hence, we can conclude that the model is a near perfect fit and does not suffer from overfitting.

From Figure 13, we can see clearly that the ROC (receiver operating characteristic curve) and the area under the ROC curve (AUC) for the 10 classes are near to 1, which signifies that our classifier model can nearly distinguish between all the positive and negative class points correctly.

From Figure 14, we can see clearly that the ROC and AUC for the 10 classes are near to 1, which signifies that our classifier model can nearly distinguish between all the positive and negative class points correctly.

The confusion matrix of Conv-LSTM GRU method on our dataset is shown in Figure 15.

The confusion matrix of LRCN method on our dataset is shown in Figure 16.