HAR is an emerging area of artificial intelligence that recognises and classifies human actions utilising data from sensors, cameras, and wearable devices. Its applications include healthcare, Smart homes, Surveillance, Violent activity detection [1], Academia, driver behaviors and fitness monitoring. HAR systems analyse motion patterns to detect different types of activities. HAR activities are classified as either normal or abnormal. Normal activities are those which are described as regular or daily, for instance, walking, sitting, standing, running, climbing stairs, and so on. Such activities illustrate standard human motion patterns and are widely utilized to train models for behavior understanding. On the other hand, unusual events are unpredictable or anomalous events, such as falling, fighting, elements of sudden collapse, or even some strange dance movements, which are culturally consistent and could be regarded as anomalies but of a different nature.

In surveillance and security, HAR also helps identify suspicious behaviors and enhance security protocols. It also allows for more natural interfaces in interactive media and human-computer interaction. Growing demands for automation, personalized services, and real-time monitoring have raised HAR as a key research area that challenges more accuracy, adaptability, and efficiency.

In HAR, accurately identifying and categorising complex human activities, such as fights, snatching, and running, remains difficult, especially when each activity has multiple subtypes with significant intra-class variation. Most existing HAR systems struggle to distinguish visually similar actions, handle occlusions, and remain robust in dynamic, real-world environments with multiple viewpoints.

Deep learning and pre-trained methods have shown significant success in feature extraction, and Convolutional Neural Networks (CNNs) have been effective in acquiring spatial patterns for activity identification [2]. Beyond the limits of conventional models, CNNs are a powerful way for analyzing spatial and image data. Additionally, CNNs preserve spatial relationships, ensuring coherence even when objects are repositioned. It is better suited for image classification and object detection tasks because it can handle noise and slight data variations, learn deeper features in deeper networks, and perform well on large and complex datasets. Detection of these challenging activities creates problems because the discussed activities are heterogeneous, complex, and not simple [3].

To overcome these challenges, this study first examines the performance of CNN-based architectures and pre-trained models separately, then presents an average-weighted fusion method that capitalizes on their respective advantages. The fusion strategy attains higher recognition accuracy by exploiting the spatial learning capabilities of CNNs and enhanced feature extraction of pre-trained models.

The main objective of this research is to develop a deep learning-based framework for classifying human activities using a labelled dataset comprising three primary activity classes: fighting, snatching, and running, each with five distinct subtypes. This research aims to create a robust human activity detection system capable of detecting safety-critical events. It explores and fine-tunes advanced deep learning models to assess their performance in precisely classifying three different classes, using five variations of each class. The majority of the literature focuses on relatively simple actions such as walking, jogging, clapping, and wrestling, which limits their applicability to real-world, high-risk circumstances. There are models specifically designed to identify complex, safety-critical actions, such as violent or abnormal behaviour (e.g., cellphone theft or fighting), which are very important for surveillance and security applications.

To address this limitation, we selected a benchmark dataset that encompasses cellphone snatching, fighting, and running activities. Deep learning was also used with CNNs and three pre-trained models, which were combined using a Weighted Average Ensemble to achieve robust and accurate classification of critical actions. The results of experiments demonstrate that the proposed fusion scheme has better performance than stand-alone models, implying the need to consider a holistic view in improving HAR performance. The study focuses on abnormal and high-risk activities (e.g., cellphone snatching, fighting, and running) that are rarely addressed in HAR research. Additionally, it utilized the Weighted Average Ensemble of CNNs with three pre-trained models, which improves generalization and reliability for surveillance-oriented HAR.

The main research contributions include:

We investigated many cutting-edge pre-trained deep learning models for classifying human activities, evaluating their strengths and limitations in extracting useful information.

The remaining sections of the paper are organized as follows: Section 2 presents a literature review; Section 3 briefly describes the proposed methodology; Section 4 discusses the results and their implications; Section 5 highlights the discussion and finally, Section 6 provides the conclusion.

This section examines current studies on human activity recognition that employed machine learning, deep learning, and Internet of Things (IoT)-based techniques. HAR uses information gathered from sensors, cameras and other sources to categorise and recognise human activities. This section summarizes some of the significant recent contributions to HAR across different approaches, datasets, results, and challenges. The work aimed to develop the model and report on its accuracy using a test population of individuals aged 19 to 48 years. Various procedures were applied to clean and process the data, as well as to identify different patterns of human activities. In terms of model learning, it performed better than other types of models, including machine learning models, with the highest accuracy rate of 88.6% among machine learning models. It was also the best deep learning model, achieving an accuracy of 84.5% on Gated Recurrent Unit (GRU) [4].

The purpose of this work [5] was to develop an activity recognition method, based on deep learning, by extracting features from the raw input data. Two Hybrid Learning Algorithms (HLA) are proposed to identify sequential and spatial patterns using CNN and Recurrent Neural Networks (RNNs). A hybrid optimisation method, combining the Whale Optimisation Algorithm and the Grey Wolf Optimiser, enhanced feature selection. This evolved the entire optimization process.

In this work, Ref. [6] proposed a dynamic method for HAR that combines a deep bidirectional long-term memory and the pre-trained feature extraction technique. Initially, Convolutional Neural Network models, particularly MobileNetV2, were used to extract deep-level characteristics from video frames. These extracted features were subsequently analyzed using an optimized Deep BiLSTM network which effectively captured dependencies and increased prediction accuracy. During testing phase, an iterative fine-tuning approach was employed to update high-level parameters and enhance flexibility in response to various conditions. The model’s performance was assessed using three benchmark datasets: UCF Sport [7], UCF11 [8], and JHMDB [9], with accuracies of 76.30%, 93.3%, and 99.20%, respectively.

This work [10] used XGBoost to build an activity recognition system. Initially, a Hue, Saturation, Value (HSV) colour transformation was used to improve video frame clarity, followed by noise reduction. Silhouettes were extracted using multiple object tracking (MOT) and Video Inference for Body Estimation (VIBE) methods. For feature extraction, Textone maps and Features from the Accelerated Segment Test (FAST) [11] were used. Independent Component Analysis (ICA) was then used to determine the most informative components for feature discrimination. Finally, the composed features were loaded into XGBoost for categorization as relevant human activities. Experiments conducted using the Stony Brook University (SBU) [12] Interaction dataset achieved a recognition rate of 91%. The AI-based behavior biometrics framework used a dynamic attention fusion unit and a temporal-spatial fusion method to improve the detection of human activity in surveillance systems. The system records temporal, geographical, and behavioral dependencies in video data streams and extracts important features employing a lightweight EfficientNetB0 backbone. The system outperformed state-of-the-art techniques, showing high accuracies of 80.342%, 98.987%, 98.734%, and 98.927%, respectively, using four publicly accessible datasets (HMDB51 [13], UCF50 [14], UCF101 [15], and YouTube Action [8]).

Some artificial intelligence models have been created to recognise human activity but have produced poor results in HAR in real-world, long-term settings. Because they have extracted insufficient temporal and spatial features [16]. RNN and the CNN model are applied to examine the best efficiency of daily activities, such as walking, and to attain the best figure of accuracy [17]. One feature of the Wi-Fi signal is Channel State Information (CSI), which identifies various human activities. Human activity data was collected using a Raspberry Pi 4 and applied techniques, including 1D-CNN, LSTM, and Bi-directional LSTM, to achieve 95% accuracy [18]. Similarly, 100% success is achieved in recognising human posture and action using a CNN model, as compared to the Microsoft Research (MSR) and Kinect Activity Recognition Dataset (KARD) [19]. A novel acceleration-based HAR method is employed in conjunction with the CNN model, utilising a large dataset of 31,688 samples. Here, eight distinctive human activities were examined and achieved 93.8% accuracy [20]. Furthermore, an advanced system for human image tracking (HIT) was developed to detect both simple and complex user movements. A smart camera was used to collect the dataset for the model of a region-based CNN. As a result, 98.53% accuracy was achieved through the proposed model [21]. Another deep learning model, CNN-LSTM, was identified to recognise normal individuals and straightforward actions, including driving and eating. Here, CNN was used to extract features from sensor data, and an LSTM was used to obtain long-term dependencies between two events. The proposed model achieved a maximum recognition rate of 95.8% [22]. In some places, a combination of three models was used, including LSTM, CNN, and CNN-LSTM, to organize human activities. These models were trained on a dataset comprising six classes of activities, collected from 36 people. Activities included walking, sitting, jogging, standing, and climbing stairs. The created model was evaluated using the TensorFlow framework, and the testing accuracy was 94.51% for the CNN, 96.61% for the LSTM, and 97.76% for the CNN-LSTM. The accuracy was checked by a precision of 97.75%, a recall of 97.77%, F1-Measure 97.76%, and an area under the curve of 100% for the CNN-LSTM model [23].

The model’s effectiveness in personalised and adaptive learning environments was validated through experimental results on datasets, including the recently created Human Activities and Postural Transitions (HAPT) dataset [24], which achieved better classification accuracy (97.84% for transitional activities and 99.04% for dynamic activities) than state-of-the-art methods.

The research introduced an ensemble deep learning approach for capturing temporal correlations in sensory data. We proposed a hybrid LSTM-GRU architecture with dropout and batch normalization and evaluated it on the UCI-HAR and Wireless Sensor Data Mining (WISDM) [25] datasets.

The method was developed to promote a culture of fluid, scalable, and ubiquitous learning The results on HAPT outperformed all the state-of-the-art methods (97.84% for transitional activities and 99.04% for dynamic activities), which demonstrates its effectiveness in the personalized and adaptively learning applications [26]. The authors proposed the method by using the motion capturing system with marker-based motion capturing system (MBMCS) and marker-less motion capturing system (MLMCS), and compared them for the analysis of human gait of a person and activity recognition. Results demonstrate that MBMCS performs up to 99.3% accuracy for person recognition and 98.1% for activity recognition using K-Nearest Neighbors, while MLMCS provides a cost-effective solution with comparable performance and the potential for further improvement by advanced feature extraction [27]. The existing work has been described in Table 1.

The proposed methodology has been discussed in this section. The diagram of the proposed framework is shown in Fig. 1. The basic steps of the proposed method are given below: (a)

Data set selection

The study results were evaluated using three DL models: ResNet50, EfficientNet, and the proposed CNN. Then, these models were fused by averaging for further performance enhancement. ResNet50 with its residual learning framework had excellent feature extraction, however, it struggled to capture more complicated activity sequences. EfficientNet is also a compound scaling method which compares to ResNet50 and is computationally efficient while achieving better accuracy performance. The proposed CNN model for human activity recognition through salient activity features has shown promising results which are tailored accordingly. However, as the models were complementary, some variations existed at the level of the different activity classes, as the unique strengths and weaknesses of each model applied. An average-weighted fusion scheme was simulated to address such a concern. ResNet50, EfficientNet and Prose CNN predictions are weighted based on the performances of the individual models. Benefitting from complementary of three models, the integrated approach improves the classification accuracy, precision, sensitivity as well as the F1 value significantly. In the case of human activity recognition, such an ensemble method has been proven to improve accuracy and to reduce classification errors and over-fitting. Our experiments demonstrate that fusion strategy is a simple yet effective approach to boost the accuracy of activity detection tasks by mitigating the limitations of single models.

The proposed weighted average ensemble framework for HAR combines various deep learning architectures to improve classification accuracy and model robustness. CNN can help to detect low-level spatial features, ResNet handles this by using skip connections to learn deeper features, whereas EfficientNet provides optimized scaling to balance accuracy and computational efficiency. The weighted average ensemble aggregates predictions and assigns importance based on model performance, combining the individual models’ strengths. Comparisons of the results using ResNet, EfficientNet, CNN, and the Proposed model with Weighted Average Ensemble have been shown in Fig. 7.

We aimed to effectively classify and recognise these complex activity patterns using deep learning models. To this end, we evaluated several state-of-the-art convolutional neural networks, including ResNet, EfficientNet, a custom-designed CNN model, and a proposed model integrated with a weighted-average ensemble approach. The experimental results demonstrated that while standard pre-trained models performed reasonably well, the proposed CNN model showed improved classification performance by capturing more discriminative spatial features. In addition, combining several models with the weighted-average ensemble method achieved the best accuracy and stability for all activity types and subtypes. In summary, the proposed ensemble model achieved superior performance than the other baseline models, suggesting that it is able to effectively capture the complexity and diversity of real-world activity recognition. These results demonstrate that ensemble-based deep learning methods are very promising to enhance the performance and robustness of HAR systems, especially in such safety-critical applications as public security, crime prevention, and surveillance. The proposed method yielded encouraging results, but the speed of processing and the overall efficiency may be further enhanced. When the focus of future work will be on performing real-time experiments, considering the problem of imbalanced data, and extending the current model to predict more human activities in diverse domains. Multimodal sensor data and self-supervised learning methods can be incorporated to demonstrate an even higher level of robustness and generalization.