1 Introduction

Notable success achieved in the monitoring of human actions allows a variety of advanced multimedia applications (Singh et al. 2020a). Owing to its great importance, human activity recognition is exploited for applications like intelligent video surveillance, abnormal behavior recognition, sports, transportation, web and healthcare. It is found in literatures that computational methods are well capable of recognizing normal and abnormal activities from image and video sequences such as walking, fighting and robbing (Nigam et al. 2019; Sahoo et al. 2020; Rajagopal et al. 2020; Pillai et al. 2021; Yousef et al. 2022; Srivastava et al. 2021). Visual surveillance equipment are used for monitoring human activities and retrieving relevant information. These are further integrated to build advanced systems (Lv et al. 2021a, 2022). Depending on number of persons involved in it, human activities can be of four types—(i) actions involving one person, (ii) interactions involving two people, (iii) interactions involving a person and an object, and (iv) group interactions involving a number of persons.

Even though action recognition is a challenging issue, much work is done on it over the past decade (Binh et al. 2013; Poppe 2010; Weinland et al. 2011; Aggarwal and Ryoo 2011; Ji and Liu 2009). One of the major issues with actions performed in a real 3D environment is that cameras capture only 2D projects of real-time actions. Therefore, visual analysis of the activities performed in image plane is merely a projection of actual real-world actions. This projection is based on viewpoints and does not contain complete information about the action. As a solution to this problem, concept of exploring information obtained from multiple cameras mounted on different viewpoints is used (Ji and Liu 2009). Therefore, the mechanisms are developed by presenting a view independent analysis for multiple views (Ji and Liu 2009). Exploring information from multiple scenes enhances the accuracy of activity recognition by obtaining features from various 2D viewpoints to achieve visual consistency. Main objective of the whole scenario is to develop a reliable human activity recognition system.

We participate in the aforementioned solution and introduce an approach for multiview human activity recognition system for image sequences. The proposed framework consists of three steps:

  1. (i)

    Finding the human objects by removing the background.

  2. (ii)

    Extraction of uniform and rotation invariant characteristics of LBP.

  3. (iii)

    Identification of human activities using SVM.

We use simple frame differencing approach for background removal from input data. After background removal, we extract uniform rotation invariant LBP features. Due to rotation invariance characteristic of uniform LBP, it provides an independent analysis of human activity perceptions. This feature is categorized using a radial basis function (RBF) kernel-based SVM classifier with one versus all (OVA) structure. Use of SVM is influenced by the fact that non-sequential strategies, like SVM, are highly competitive and balanced in large-scale and continuous work data (Nigam et al. 2018). Multiclass classification is achieved using hierarchical organization of several binary classifiers.

To illustrate the effectiveness of our proposed work, experiments are performed on three benchmark and publicly available multiview human activity video datasets. These datasets are Weizmann, CASIA and INRIA Xmas Motion Acquisition Sequences (IXMAS). We evaluate the proposed method with the existing and established feature descriptor based methods. The test results of the three datasets show the efficacy of the proposed framework.

Following are the major highlights of the proposed work.

  1. (i)

    We introduce a rotation invariant human activity recognition framework.

  2. (ii)

    Multiview human activity recognition is handled with background removal.

  3. (iii)

    Uniform LBP and SVM classifier are exploited for implementation of the proposed framework.

Organization of this framework is as follows. Section 2 briefly discusses the related works. Section 3 elaborates the development of LBP features and the organization of the SVM classifier for multiple classes. Section 4 provides the implementation details of the proposed method. The results of the evaluation and discussion on the three public datasets are provided in Sect. 5. Section 6 provides conclusive remarks of the study.

2 Related works

Recognition of human activity is the process to detect human body motion patterns. Popular devices to detect human activity are sensors and cameras. Mostly, two type of activity recognition systems do exist, one is sensor-based and other is vision-based. Vision-based systems are more popular as compared to sensor-based since they provide important cues of activity recognition. Many researchers have contributed towards the reviews on human activity recognition (Saha et al. 2022). Based on these reviews, activity recognition approaches can be model-based and model-free.

Model-based activity recognition uses a pre-model to monitor human activities. These vivid 2D and 3D shape models are used to visualize people's activities. Global and local features have been combined in (Wang and Mori 2010) to implement a framework for human action recognition. This work has demonstrated that the combination of part-based model and motion features with large-scale improves the results. Instead of constructing hidden part model, the work in (Wu et al. 2014) has constructed hidden temporal models for each action class. It has focused on human action recognition in uncontrolled videos containing complex temporal structures. The work in (Lan et al. 2011) has focused on the recognition of specific actions and group activities. It has also defined a new feature called action context descriptor. This approach has demonstrated good visual results of several complex but mathematically costly tasks. Cheng et al. (2014) has developed a layered model to represent group activities at diverse granularities. New informative descriptions of the appearance of group actions have been introduced in this way. A nearest neighbour classifier and Gaussian mixture model based work has been proposed for video action recognition using motion curves in (Vrigkas et al. 2014).

However, from the analysis of the model based methods, it is observed that there is a trade-off in retrieving a detailed knowledge of the human body, and the cost of calculation and robustness. The model based methods exploit the pose and velocity vectors which may increase the computational complexity. Sometimes, major parts of body models are taken to reduce the complexity (for example hands, legs, torso etc.), still it is difficult to construct these models. Also, model based approaches need to be implemented directly and could not work in real time.

Model-free approaches overcome shortcomings of model-based approach. In model-free methods, low visibility features from area of interest are retrieved for action recognition. These methods are based on posture, global and local motions (Määttä et al. 2010). The feature based multiple view approaches obtain image data captured by multiple cameras. A two-camera based method has been implemented for multiple humans pointing in a direction (Matikainen et al. 2011). The different views of 2D pointer configurations have been used to obtain 3D pointing vectors. Five calibrated and synchronized cameras have been used in (Souvenir and Babbs 2008). R transform and manifold learning of the silhouettes have been used for view invariant activity recognition. The circular shift invariance nature of discrete Fourier transform have been exploited in (Iosifidis et al. 2010).

Data fusion has also been exploited for multiview human activity recognition (Weinland et al. 2010). It has used 3D histogram of oriented gradients (HOG) features and applied local partitioning along with hierarchical classification on it. A similar method has been implemented using view point aggregation and multiview dynamic image fusion for cross view 3D action recognition (Wang et al. 2021). It has used 3D characterization and fisher vector for representation of 3D action.

Cross-view activity recognition is an interesting topic for researchers. This is a difficult task of human activity recognition since training and testing views are different. Numerous techniques are proposed for this purpose including learning from short video clips (Vyas et al. 2020), bilayer classification model (Li et al. 2019) and unsupervised attention transfer (Ji et al. 2021).

In recent works, deep learning and transfer learning have become useful tools (Lv et al. 2021b; Singh et al. 2020b). Few deep learning-based techniques are defined in (Jan and Khan 2021; Verma and Singh 2021; Verma et al. 2020) which are very efficient to perform recognition task.

Today, dynamic texture patterns like LBP, have become an obvious choice for the recognition of the activity of a person considered as moving texture patterns. A few examples of them are (Nigam et al. 2021; Kellokumpu et al. 2010, 2011; Vili et al. 2008). However, none of these strategies uses the rotation invariant uniform LBP. Selecting such patterns reduces length of LBP histogram and improves efficacy of a classifier (Pietikäinen et al. 2011). It is widely accepted that uniform LBP is highly effective and has been used repeatedly in several other applications in addition to texture analysis (Bianconi and Fernández 2011). Although many upgraded versions of simple LBP have been introduced, many techniques still benefit from the uniform LBP. However, it is not clear that how the uniform patterns contribute to the LBP based discrimination (Lahdenoja et al. 2013). Furthermore, uniform LBP has been used successfully to obtain rotation invariance (Fernández et al. 2011). The use of uniform binary patterns with rotation invariance is advanced version when compared to its predecessors, as they provide additional integrated representation (Ojala et al. 2002). The global rotation of LBP has been achieved in (Ahonen et al. 2009) using the discrete Fourier transformation in the uniform LBP histograms bins. Rotation invariance characteristic of LBP variants has also been discussed in (Zhao et al. 2011).

From above description of human activity recognition literature, it can be inferred that uniform LBP results in better selection of human multiview activity recognition.

3 Principles and basics

This section briefly discusses two major components of the proposed method, which are the uniform rotation invariant LBP and the multiclass SVM.

3.1 Uniform and rotation invariant LBP

  • LBP

The LBP feature is built for a circular neighbourhood of radius R pixel. Intensity of P sample points is compared in the circular neighbourhood with the centre pixel in either clockwise or anticlockwise direction (see Fig. 1).

Fig. 1
figure 1

Circular neighborhoods for different (P, R) (anti-clockwise)

Full size image

This comparison determines whether the pixel value should be zero (0) or one (1). A value 0 is given if the median pixel magnitude is greater than the neighborhood pixel and a value 1 is given if the median pixel magnitude is less than neighborhood pixel. A popular option is 8 for the number of sample points in the neighborhood and 1 for radius (i.e., P = 8 and R = 1). Although, other combinations may also be used. Intensity of a sample point between two pixels is determined by the bilinear interpolation. The LBP feature of an image is denoted by LBP (Pietikäinen et al. 2011). After having extracted the LBP of a pixel, intensity value of the pixel is replaced by this LBP. This procedure could not be followed for border pixels because all of the neighbor values do not exist there. Under these considerations, feature vector of an image is given by

$$LB{P}_{P,R}(x,y)={\sum }_{P=0}^{P-1}s({g}_{P}-{g}_{c}){2}^{P}$$
(1)

In Eq. (1), (x, y) is the center pixel location, gc represent the center pixel intensity, gp represent the pixel intensity of the neighborhood and s(w) is defined as

$$s(w)=\left\{\begin{array}{c}1, w\ge 0\\ 0, w<0\end{array}\right\}$$
(2)

The feature vector \(LB{P}_{P,R}\) of an image is LBP histogram of all pixels in this image. The initial dimension of this LBP histogram is \({2}^{P}\) since each LBP may be assigned to a different bin. If an image has M regions, then total number of histograms formed in the image are M · \({2}^{P}\) or we can say that the image has a histogram whose size equals to M · \({2}^{P}\).

  • Rotation invariance

Several upgraded versions of LBP, as discussed in (Pietikäinen et al. 2011), have been developed to achieve invariance against rotation and to reduce the size of LBP histogram. For the rotated image, the gray value gp will shift along the rotation circle perimeter, hence different \(LB{P}_{P,R}\) can be calculated. To reduce the effect of rotation, an upgraded LBP including invariance against rotation is defined as

$$LB{P}_{P,R}^{ri}(x,y)=\mathit{min}\left\{ROR(LB{P}_{P,R},i)|i=\mathrm{0,1},....,R-1\right\}$$
(3)

In Eq. (3), \(ROR(LB{P}_{P,R},i)\) makes a circular bitwise right shift i times to the R-bit number \(LB{P}_{P,R}\). The \(LB{P}_{P,R}^{ri}\) feature vector can have 36 different values for R = 8, and it can have the histogram size 36 for an image region.

  • Uniform patterns

Uniform LBP is having 0, 1 or 2 circular transitions between binary value 0 and 1. Let us take few examples of uniform and non-uniform patterns. The 0–1 transitions of uniform patterns for P = 8 and R = 1 are shown in Fig. 2. In a circular neighborhood of P pixels, number of uniform patterns found is P + 1. A brief description of uniform and non-uniform patterns is shown in Table 1.

Fig. 2
figure 2

Uniform LBP (P = 8, R = 1). Hollow circle indicates 0 and solid circle indicates 1

Full size image
Table 1 Uniform and non-uniform patterns
Full size table

Formal definition of uniform LBP with rotation invariance is given by

$$LB{P}_{P,R}^{riu2}=\left\{\begin{array}{l}{\sum }_{P=0}^{P-1}s({g}_{P}-{g}_{C}), if \, {\text{U}}({\text{L}}{\text{B}}{\text{P}}_{P,R})\le 2\\ P+1, \, \, \, \, \, \, \, \, \, \, \, otherwise\end{array}\right.$$
(4)

where riu2 denotes rotation invariant uniform patterns, gc represent the center pixel intensity, gp represent the pixel intensity of the neighborhood, s(w) is defined in Eq. (2) and

$$U(LB{P}_{P,R})=\left|s({g}_{P-1}-{g}_{C})-s({g}_{0}-{g}_{C})\right|+{\sum }_{P=1}^{P-1}\left|s({g}_{P}-{g}_{C})-s({g}_{P-1}-{g}_{C})\right|$$
(5)

3.2 SVM multiclass classifier

Initially, SVM classifier was proposed for binary classification and subsequently used for multiple classification successfully.

  • SVM classifier

For a set of \(L\) points, where \({x}_{i}\) consists of \(D\) attributes and belongs to class \({y}_{i}=+1\, or\,-\,1\). Hence, the training set is

$$ \left\{ {x_{i} ,y_{i} } \right\},i = 1............L,{ }y_{i} \in \left\{ {1, - 1} \right\},x \in R^{D} $$
(6)

For linearly separable data, a hyperplane can be drawn for graphs of \({x}_{1},{x}_{2},\dots , {x}_{D}\) where \(D>2\). SVM provides the hyperplane which is closest from both classes.

The data point \({x}_{S}\) is represented by

$$ y_{s} \left( {\sum _{{m \in S}} \alpha _{m} y_{m} x_{m}\cdot x_{s}+ b} \right) = 1 $$
(7)

where \(\alpha \) is called Lagrange Multiplier satisfying \({\alpha }_{i}\ge 0 {\forall }_{i}\)and \(S\) represents the set of indices of support vectors. \(S\) is obtained by indices \(i\) where \({\alpha }_{i}\ge 0\). The support vectors in S are

$$b=\frac{1}{{N}_{s}}{\sum }_{s\in S}({y}_{s}-{\sum }_{m\in s}{\alpha }_{m}{y}_{m}{x}_{m}\cdot {x}_{s})$$
(8)

  • One Versus All (OVA) architecture

The OVA architecture is initially used to implement multiclass classification with SVM. The idea behind it is that all classes are divided into two categories. One class is included in positive class at a time and all other classes are included in the negative one. This process is repeatedly used for each and every class. Hence, classifier is applied same number of times as the classes are. Architecture of this classifier is shown in Fig. 3.

Fig. 3
figure 3

SVM multiclass classifier with OVA architecture

Full size image

4 The proposed framework

This section presents proposed framework for human activity recognition system. Figure 4 represents the proposed framework which has been separately discussed in this section.

Fig. 4
figure 4

Block diagram of the proposed framework

Full size image

4.1 Input video

It is a sequence of frames which is used for training purpose. The video can have number of frames from \(1, 2,\dots , n\). This video is represented as

$$V={\sum }_{i=1}^{n}{F}_{i}$$
(9)

where \(F\) denotes a particular frame and \(i\) is the number of frames which lies among \(1, 2,\dots , n\).

4.2 Preprocessing

The preprocessing step is performed in order to reduce color and size variations and to have all video frames in a consistent format. Therefore, the normalized video is

$$\left|V\right|={\sum }_{i=1}^{n}\left|{F}_{i}\right|$$
(10)

4.3 Background subtraction

Foreground object detection is an important step in activity recognition process. Background subtraction is a simple yet efficient technique for this purpose. Therefore, in the proposed framework, foreground object detection is performed using background subtraction approach for capturing the human activities only. Frame differencing is a popular method for this. However, threshold selection is an important step in execution of background subtraction. Few representative results of background subtraction are shown in Fig. 5. A general algorithm for frame differencing based background subtraction is provided below.

figure a
Fig. 5
figure 5

Original and corresponding background subtracted images

Full size image

4.4 Feature vector extraction

The next step is to extract the features from the background subtracted video \(\left|{V}_{bs}\right|\). This motion feature is computed as following equations

$$LB{P}_{P,R}^{riu2}=\left\{\begin{array}{l}{\sum }_{P=0}^{P-1}s({g}_{P}-{g}_{C}), \, if\text{ U}({\text{LB}}{\text{P}}_{P,R})\le 2\\ P+1, \, \, \, \, \, \, \, \, \, \, \, \, \, otherwise\end{array}\right.$$
(11)

where \(s(w)=\left\{\begin{array}{c}1, w\ge 0\\ 0, w<0\end{array}\right\}\), (12)

$$U(LB{P}_{P,R})=\left|s({g}_{P-1}-{g}_{C})-s({g}_{0}-{g}_{C})\right|+{\sum }_{P=1}^{P-1}\left|s({g}_{P}-{g}_{C})-s({g}_{P-1}-{g}_{C})\right|$$
(13)

\({g}_{c}\)= center pixel of \(\left|{V}_{bs}\right|\) and \({g}_{p}\)= neighborhood pixel of \(\left|{V}_{bs}\right|\).

4.5 Recognition of activities

SVM is an efficient classifier that has been extensively used in literature for recognition task. We have used SVM in the proposed framework since it produced high accuracy and performed faster predictions in comparison to other existing classifiers. After feature extraction step, activity recognition step is performed by training and testing of SVM multiclass classifier. Among different kernel functions, RBF kernel is selected owing to its better accuracy. The input video V belongs to specific activity subject to corresponding class

$$V\in \left\{\begin{array}{c}activit{y}_{1} \, \, \, iff \, \, \, V\in clas{s}_{1}\\ activit{y}_{2} \, \, \, iff \, \, \, V\in clas{s}_{2}\\ \dots \dots ........................\\ activit{y}_{n} \, \, \, iff \, \, \, V\in clas{s}_{n}\end{array}\right.$$
(14)

5 Experimentation and results

This section performs experimentation on 3 benchmark datasets—the Weizmann viewpoint dataset (Gorelick et al. 2007), the CASIA dataset (Wang et al. 2007) and the IXMAS dataset (Weinland et al. 2006). First input video is supplied to the pipeline. After that preprocessing, background subtraction and feature extraction steps have been executed followed by classification. Both qualitative and objective evaluation of the proposed framework is performed and it is tested in comparison to moment based method (Binh et al. 2013), moment invariant based method (Nigam and Khare 2016), center symmetric LBP based method (Bianconi and Fernández 2011), and GIST feature based method (Vyas et al. 2020).

5.1 Case study 1

This experiment illustrates the efficacy of the proposed framework for various rotations of an activity. For this purpose, we have selected Weizmann viewpoint dataset (Gorelick et al. 2007). It contains 10 videos showing “walking” activity captured from cameras placed at different viewpoints. These viewpoints vary between 0° and 81° angle relative to the image plane with a difference of 9°. With the gradual increase in the angle of the videos, the videos contain significant changes in scale. The videos are of frame size 180 × 144, image type is ‘TrueColor’ and no video compression is used. We have selected 40 frames from each video.

We have performed background subtraction followed by feature vector computation and classification. Visual background subtraction results are shown in Fig. 6 and recognition results are shown in Fig. 7.

Fig. 6
figure 6

Background subtraction results

Full size image
Fig. 7
figure 7

Visual recognition results

Full size image

The angle changes from 0° to 27°, 54°, 63°, 72°, and 81°. From Fig. 6, we get background subtracted images of the activity “walking” being performed at different rotation angles. In this case, we get foreground object segmented correctly. Not much noise is present in the images and one can easily recognize the foreground object in these videos. From Fig. 7, we get visual recognition results for activity “walking” at different rotation angles. From these results, we get correct results for recognizing the activity whether it is being performed at different rotation angles. The angle of rotation has been recognized correctly for 10 different viewing angles. Hence, one can get correct background subtracted images and visual results.

Now, we have shown quantitative results and comparatively evaluated in terms of confusion matrix and recognition accuracy. The compared methods are moment based method (Binh et al. 2013), moment invariant based method (Nigam and Khare 2016), center symmetric LBP based method (Bianconi and Fernández 2011), and GIST feature based method (Vyas et al. 2020). This section shows quantitative results for Weizmann viewpoint dataset (Gorelick et al. 2007). Confusion matrices and recognition accuracy have been shown in Tables 2, 3, 4, 5 and 6. From these tables, it is observed that the our method performs better than other methods.

Table 2 Confusion matrix for moment based method (Binh et al. 2013)
Full size table
Table 3 Confusion matrix for inv-mom based method (Nigam and Khare 2016)
Full size table
Table 4 Confusion matrix for CSLBP based method (Bianconi and Fernández 2011)
Full size table
Table 5 Confusion matrix for GIST based method (Vyas et al. 2020)
Full size table
Table 6 Confusion matrix for the proposed method
Full size table

Recognition accuracy of different methods for WEIZMANN dataset is shown in Fig. 8. From this table, we can observe that recognition accuracy of moment-based method is 10%, invariant moment-based method is 22.7%, CSLBP method is 32.9%, GIST based method is 33.2% and for the proposed method, it is 100%. From this table, we can conclude that the proposed framework outperforms other methods.

Fig. 8
figure 8

Recognition accuracy of different methods

Full size image

5.2 Case study 2

This section demonstrates performance of the proposed framework for CASIA action recognition dataset (Wang et al. 2007). Five most common interactions have been selected which are fight, overtake, rob, follow, and meet and part. All videos have been captured from three different viewpoints: angular, horizontal and top down. The activities have been renamed as fight: angular view, fight: horizontal view, fight: topdown view, overtake: angular view, overtake: horizontal view, overtake: topdown view, rob: angular view, rob: horizontal view, rob: topdown view, follow: angular view, follow: horizontal view, follow: topdown view, meet and part: angular view, meet and part: horizontal view, and meet and part: topdown view. Videos have been compressed using HUFFYUV compression technique in AVI video format. Duration varies between 5–30 s for different activities. Visual background subtraction and recognition results are shown in Figs. 9 and 10, respectively.

Fig. 9
figure 9

Background subtraction results

Full size image
Fig. 10
figure 10

Visual recognition results

Full size image

Figures 9 and 10 show that we get correct background subtracted frames and visual results. The activities have been recorded using three different viewing angles. The background removal step of the proposed method provides better results which lead to correct object identification. As a result, we have better recognition of human activities and the proposed method is able to identify activities for multiple views. The proposed method also identifies suspicious activities which are fight, overtake, rob, follow, and meet and part.

Now, CASIA dataset have been analyzed and presented (Wang et al. 2007). The activities are renamed as; Fight: angular view as A1, fight: horizontal view as A2, fight: topdown view as A3, overtake: angular view as B1, overtake: horizontal view as B2, overtake: topdown view as B3, rob: angular view as C1, rob: horizontal view as C2, rob: topdown view as C3, followalways: angular view as D1, followalways: horizontal view as D2, followalways: topdown view as D3, meetapart: angular view as E1, meetapart: horizontal view as E2 and meetapart: topdown view as E3. The results have been shown in Tables 7, 8, 9, 10 and 11. Although, CSLBP based method (Bianconi and Fernández 2011) shows a high recognition rate but proposed framework performs better than this method also.

Table 7 Confusion matrix for moment based method (Binh et al. 2013)
Full size table
Table 8 Confusion matrix for inv-mom based method (Nigam and Khare 2016)
Full size table
Table 9 Confusion matrix for CSLBP based method (Bianconi and Fernández 2011)
Full size table
Table 10 Confusion matrix for GIST based method (Vyas et al. 2020)
Full size table
Table 11 Confusion matrix for the proposed method
Full size table

Recognition accuracy of different methods for CASIA dataset is shown in Fig. 11. From this figure, we can observe that recognition accuracy of moment-based method is 6.7%, invariant moment-based method is 10.3%, CSLBP method is 83.7%, GIST based method is 26.7% and for the proposed method, it is 90.7%. Although, performance of the CSLBP is quite comparable to the proposed one but when we consider overall accuracy of all three cases then we can conclude that the proposed framework is better than the CSLBP based method.

Fig. 11
figure 11

Recognition accuracy of different methods

Full size image

5.3 Case study 3

Now, we have selected INRIA IXMAS multiview activity dataset (Weinland et al. 2006). This includes 13 daily-life activities performed by 11 actors 3 times each. These activities are kick, punch, turn-around, check watch, cross arms, scratch head, sit down, get up, walk, wave, point, pick up, throw (overhead), throw (from bottom up) and nothing captured from five calibrated cameras. Qualitative background subtraction and recognition results are shown in Figs. 12 and 13. From Fig. 12, we see that the foreground objects have been segmented properly. In Fig. 13, we show the obtained visual results for the proposed method.

Fig. 12
figure 12

Background subtraction results for Kick activity

Full size image
Fig. 13
figure 13

Visual recognition results

Full size image

Now, objective evaluation results are shown for IXMAS dataset (Weinland et al. 2006). Objective evaluation results have been listed in Tables 12, 13, 14, 15 and 16 which infer that the proposed framework is better than other methods.

Table 12 Confusion matrix for moment based method (Srivastava et al. 2021)
Full size table
Table 13 Confusion matrix for inv-mom based method (Jan and Khan 2021)
Full size table
Table 14 Confusion matrix for CSLBP based method (Bianconi and Fernández 2011)
Full size table
Table 15 Confusion matrix for GIST based method (Wang et al. 2021)
Full size table
Table 16 Confusion matrix for the proposed method
Full size table

Recognition accuracy of different methods for IXMAS dataset is shown in Fig. 14. From this figure, we can observe that recognition accuracy of moment-based method is 6.7%, invariant moment-based method is 11.2%, CSLBP method is 83.7%, GIST based method is 100% and for the proposed method, it is 90.7%. Although, performance of the GIST based method (Wang et al. 2021) is quite comparable to the proposed one but when we consider overall accuracy of all three cases then we can conclude that the proposed framework is better than the GIST based method.

Fig. 14
figure 14

Recognition accuracy of different methods

Full size image

In addition to these promising results, these experiments have some limitations also. All these experiments have been conducted on 70% training and 30% testing datasets. Number of samples in these sets can affect overall accuracy of the framework. Also, these experiments have been performed for training and testing sets from the same dataset. Hence, cross-dataset selection may also affect the accuracy of the method.

6 Conclusions

This study proposed an activity recognition system which is suitable for multiple views. Overall framework of the system consists of background subtraction, feature extraction and activity recognition. Background subtraction is used to capture the object only. Then, uniform and rotation invariant LBP descriptor is computed. Multiclass SVM classifier has been applied for multiclass recognition of different activities. The framework is tested on three multiview human activity video datasets: Weizmann viewpoint dataset (Gorelick et al. 2007), CASIA action recognition dataset (Wang et al. 2007) and IXMAS dataset (Weinland et al. 2006) and compared with (Binh et al. 2013; Vyas et al. 2020; Nigam and Khare 2016; Bianconi and Fernández 2011). This comparison shows that the proposed method outperforms other feature descriptor based methods. This work can be extended for the development of context aware activity recognition which may include multiview and cross-view 2D and 3D human activities. Also, the fusion of existing feature descriptors with machine and deep learning can be done for better representation and recognition of multiview human activities.