Introduction

Machine learning overlaps with many science fields. One of these fields is computer vision [1]. Machine learning methods provide computer vision with techniques that offer a great leap in many of its applications, for instance, object recognition, tracking, classification, 3D modeling, and many other applications. These applications have witnessed great improvement with the emergence of new machine learning techniques. One of these applications is HAR. HAR is the basis for many other computer vision applications, for example augmented reality, robotics, automatic monitoring for medical, industrial, or surveillance purposes, etc. [2]. HAR aims at recognizing an action from data acquired by sensors, images, videos, etc. Proposed HAR systems face many limitations that narrow their applicability on domain-specific applications: limitations like (i) lighting variation, (ii) camera perspective change, (iii) scale variation, (iv) occlusion resulting from same human body parts or even from surrounding objects, (v) background confusion by undefined static or moving objects, and (vi) class confusion whether from intra-class similarities or inter-class variation [3]. Also another limitation that may rise with training of domain-specific applications is the lack of descriptive datasets. Also, proposed HAR systems face many challenges such as the following: (i) Reliability: HAR systems are required to make continuous real-time decisions based on what the system yields in surveillance applications. (ii) Social acceptance: as in smart home assistant devices, these devices closely monitor users behaviors at home, so this invades the users privacy, which raises the disclosure challenge. (iii) Disclosure: the data is required to be processed on the same device and not to be shared with any other parties for processing or advertisement. Due to the multidisciplinary nature of HAR [2], HAR requires continuous examination and exploration of new solutions to tackle these limitations and challenges.

To overcome the limitations and challenges of solving HAR problems, new techniques must be investigated. One of the most successful and recent technique is ConvLSTM [4]; ConvLSTM is a convolution LSTM layer used in many computer vision applications that require processing of spatiotemporal data. ConvLSTM uses convolution operation on an LSTM unit to perform input-to-state transformation or state-to-state transformation. ConvLSTM accepts a set of data over time to perform some prediction task. ConvLSTM has been widely used in many computer vision applications, for example, object tracking [5], scene segmentation [6], activity recognition [7], and video enhancement, like video rain removal [8]. In this paper, we propose a modification to conventional ConvLSTM layer, inspired by [9]. The novelty of our approach lies in the following points:

  1. 1

    The introduction of residual inception convolutional recurrent layer, ResIncConvLSTM, a variation of conventional ConvLSTM layer that incorporates both the concepts of residual and inception with ConvLSTM layer

  2. 2

    Solving Human activity recognition problem by designing an architecture using the newly introduced ResIncConvLSTM layer

Our proposed approach is found to outperform state of the art architecture [4] by 7% when tested on KTH dataset [10] and 11% when tested on Weizmann dataset [11]. We trained and tested our approach on a subset of UCF Sports Action dataset [12, 13] and it is found to outperform state of the art architecture [4] by 21%. Also, our proposed architecture shows efficiency improvement compared to some state-of-the-art architectures. The fact that the effectiveness of our proposed architecture is better than both the ConvLSTM baseline architecture and some state-of-the-art architectures promises better results by replacing some of the conventional ConvLSTM layers with the proposed ResIncConvLSTM layer in deeper ConvLSTM-based architectures.

Related work

HAR is a core component in various computer vision applications, so an extensive amount of research is done to tackle HAR problem. In this section, we briefly review the most remarkable and related work done to solve HAR problem. And since in this paper, we only examine visual data, we are only concerned with vision-based HAR methods. HAR approaches can be classified according to the following perspectives.

Feature extraction process

One perspective to classify HAR approaches is the process by which the features are extracted. The process of extracting features is either manually crafting features, or by using classical machine learning for automatically learning these features. Manually crafting features can either use the external seen features of the object just like body parts or motion [14–16] or a hybrid between these features [17, 18]. Many feature extraction-based methods are recently used to solve HAR. Nazir et al. [19] proposed a novel feature representation, 3D Harris space-time interest point detector and 3D Scale-Invariant Feature Transform (3DSIFT) descriptor. The features are then extracted and arranged for each frame sequence using Bag-of-features (BOF) method. Then a multiclass support vector machine (SVM) model is trained and tested using the extracted BOF. Nadeem et al. [20] proposed a HAR method based on training an artificial neural network with multidimensional features. These multidimensional features are estimated for twelve different body parts using body models. On the other hand, automatic learning of features can be done using a non-deep learning approaches like Bayesian network [21] or dictionary learning [22], or using deep learning. CNN [23], RNN [24], ConvLSTM, and CNN-RNN [25, 26] are examples of deep learning approaches. Since the introduction of ConvLSTM layer, many ConvLSTM-based architectures are introduced to solve HAR problem. In 2018, Yuki et al. [27] proposed dual-ConvLSTM to extract global and local features and use them in the recognition process. In 2019, Majd and Safabakhsh [7] introduced motion-aware ConvLSTM architecture that captures not only spatial and temporal correlations but also motion correlations between successive frames. In 2020, Kwon et al. [28] introduced hierarchical deep ConvLSTM architecture to capture different complex features.

Feature representation

One perspective to classify HAR approaches is how features are represented. Features can be arranged to preserve spatial, temporal, color, or dimensional information. The features can be flattened as in BOF method losing its spatial information. Aly and Sayed [29] proposed a HAR method that exploits both global and local features to differentiate between similar actions, like walking and running. The method first extracts from the input sequence of frames both global and local Zernike Moment features. Theses features are then combined and represented using BOF method. Then multiclass SVM algorithm is trained to recognize different actions. Another example of feature representation is silhouette frames. The input undergo silhouette extraction. Silhouette image or frame is an image that shows only the outline or the shape of the object of interest. The most important characteristic in silhouette representation is recognizing moving objects. Ramya and Rajeswari proposed, in [30], a HAR method based on silhouette frames. The method consists of three consecutive stages. First, the method preforms background subtraction to extract silhouette frames. Second, distance transform based features and entropy features are extracted from the silhouette frames. Third, a neural network is trained using these extracted features to recognize various actions. Another feature representation is bit maps, for example, features can be arranged to be a bit map of the moving object of interest. In [31], the authors proposed a solution to HAR problem using 3D CNN network receiving as an input a 3D motion cuboid. Input binarization is also one type of feature representation. Another famous feature representation is training the model with a sequence of sampled frames, where each pixel in each frame is considered a feature value. [23, 24, 32, 33] are examples for approaches that use a sequence of sampled frames as the input features.

Supervision Level

Another perspective to classify HAR approaches is whether the approach is supervised, unsupervised, or semi-supervised learning. Supervised learning approaches relies on training the model using massive well-labeled data. [34, 35] are examples of HAR supervised approaches. In [34], Han et al. proposed a two stream CNN model to perform action recognition. The authors proposed an augmentation strategy to overcome the overfitting problem that arises from the absence of massive labeled datasets. The proposed augmentation strategy is based on remodeling the dataset using a transfer learning model. In [35], Zhang et al. proposed a feature extraction approach that decouples spatial and temporal features. The proposed approach is based on a dual channel feedforward network to extract static spatial features from a single frame and dynamic temporal features from consecutive frame differences. Both decoupled spatial and temporal features are then fed to a multiclass SVM model for action recognition. On the other hand, unsupervised learning do not need labeled data. Discrimintive features are learned by the model from unlabeled data. [36–42] are examples of work done to solve HAR problem based on unsupervised learning. Abdelbaky and Aly [36–39] presented several solutions to HAR based on unsupervised deep convolution network PCANet, a simple PCANet [36], and an extended solution in which spatiotemporal features are learned from three orthogonal planes (TOP), PCANet-TOP [38]. In [40, 41], Rodriguez et al. solved HAR problem using one-shot learning in which a class representation is built from a few or a single training sequence. The authors proposed a model based on Simplex Hidden Markov Model (SHMM) and an optimized Fast Simplex Hidden Markov Model (Fast-SHMM). In [42], Haddad et al. presented a method to solve HAR problem using one-shot learning model using Gunner Farneback’s dense optical flow (GF-OF), Gaussian mixture models, and information divergence. Semi-supervised learning is a hybrid approach that benefits from the ability of supervised learning to learn features and the ability of unsupervised learning to learn hidden non-visual patterns. In semi-supervised learning, the training is done with partially labeled data. Also, not all the classes are essentially known. [43, 44] are examples of work done to solve HAR problem based on semi-supervised learning approach.

Essential background

In this paper, a new layer is proposed by integrating residual and inception concepts into ConvLSTM layer. In this section, a brief introduction about ConvLSTM, residual, and inception architectures are explained.

ConvLSTM architecture

ConvLSTM is first introduced in [4]. ConvLSTM overcomes the shortcoming of fully connected LSTM in handling spatial data. Fully connected LSTM handles temporal correlation leaving out encoding spatial data. ConvLSTM addresses this problem by applying convolution to input-to-state and state-to-state transformations. Figure 1 illustrates ConvLSTM unfolded structure and its main tensors, the inputs X1,...,Xt, cell current states C1,..., Ct, and hidden states H1,..., Ht, which are also considered cell output.

Fig. 1
figure 1

ConvLSTM unfolded structure

Full size image

(1)–(5) are the fundamental equations of ConvLSTM. Gates it,ft,ot are 3D tensors. “ σ” is nonlinear activation function, “ ∗” is convolution operator, and “ ∘” is the Hadamard product. The transformation from one state to another occurs as illustrated in (1)–(5). The next value of each cell in the grid is determined by both the input and the current value of the neighboring cells. This can be achieved by applying convolution to input-to-state and state-to-state transformations.

$$ i_{t} = \sigma (W_{xi} \ast X_{t}+ W_{hi} \ast H_{t-1} + W_{ci} \circ C_{t-1} + b_{i}) $$
(1)
$$ f_{t} = \sigma (W_{xf} \ast X_{t} + W_{hf} \ast H_{t-1} + W_{cf} \circ C_{t-1} + b_{f}) $$
(2)
$$ C_{t} = f_{t} \circ C_{t-1} + i_{t} \circ tanh(W_{xc} \ast X_{t} + W_{hc} \ast H_{t-1} + b_{c}) $$
(3)
$$ o_{t} = \sigma (W_{xo} \ast X_{t}+ W_{ho} \ast H_{t-1} + W_{co} \circ C_{t} + b_{o}) $$
(4)
$$ H_{t} = o_{t} \circ tanh(C_{t}) $$
(5)

Padding is used to make sure that both hidden states and inputs have the same dimensions. Padding of hidden states on the borders is viewed as using the state of the outside world in calculation. Hidden states are initialized with zeros to indicate total ambiguity of the future.

ConvLSTM-based architecture is first applied on next frame prediction using artificially generated moving MNIST [45] dataset in [4]. Since the introduction of ConvLSTM layer in 2015, many modifications were proposed to improve the layer architecture. The same exact year, Liang and Hu [46] introduced simple recurrent convolutional unit, RCNN, and applied it on objet recognition using SVHN [47], CIFAR-100 [48], and MNIST [49] datasets. In 2017, Alom et al. [50] proposed inception recurrent convolutional unit, IRCNN. In 2018, Wei et al. [51] introduced residual ConvLSTM and applied it on tweet count prediction. In 2020, Alom et al. [9] proposed residual inception recurrent unit, IRRCNN. IRCNN and IRRCNN are tested against CIFAR-100.

Inception architecture

Inception architecture is first introduced in [52]. Inception module is designed as shown in Fig. 2. It combines output from 3 different layers (1  × 1 conv, 3  × 3 conv, and 5  × 5 conv) and concatenates them to form the input to the next layer. Different filter sizes help detecting features that may come in different sizes. The 1  × 1 conv that precedes 3  × 3 conv and 5  × 5 is used to compute reductions before the computationally expensive layers 3  × 3 conv and 5  × 5 conv. Inception modules are recommended to be used in higher layers to extract complex features, while using conventional convolutional layers as lower layers.

Fig. 2
figure 2

Inception architecture

Full size image

Residual architecture

Residual architectures are introduced in [53]. Theoretically, as a neural network goes deeper, a more complex nonlinear function is learned. This complex function becomes able to discriminate between different classes easily. For applying this practically, sufficient dataset should be used in training this network. Training takes place by updating the current weights with values proportional to the gradient of the loss function. Also, as a neural network goes deeper a problem with gradient flow arises. The gradient becomes vanishingly small. This prevents the neural network from any further change to the network weights, so eventually the network does not learn anything new. This problem is called the vanishing gradient problem. Residual architectures are motivated by this problem. A residual neural network is based on the idea of shortcuts to skip some layers, typically two or three layers, as shown in Fig. 3. This helps with the vanishing gradient problem by reusing activation from previous layers until the adjacent layers learn and adjust their weights. This allows an alternative way for the gradient to flow.

Fig. 3
figure 3

Residual architecture

Full size image

Methods

In this section, the proposed ResIncConvLSTM layer is explained in more details. The proposed layer incorporates residual and inception concepts into the conventional ConvLSTM.

ResIncConvLSTM layer design

Figure 4 shows the design of ResIncConvLSTM layer. The input from the previous layer, X, is fed to the ResIncConvLSTM layer. ResIncConvLSTM consists of two main parts. The first part is the inception part. The inception part consists of four parallel branches:

  1. 1

    A ConvLSTM layer with kernel size 1  × 1 and number of filters fa1x1, Ha(X)

    Fig. 4
    figure 4

    Proposed ResIncConvLSTM structure labeled with data tensor dimensions and corresponding equations notation

    Full size image
  2. 2

    A ConvLSTM layer with kernel size 1  × 1 and number of filters fb1x1, followed by a ConvLSTM layer with kernel size 3  × 3 and number of filters fb3x3,Hb(X)

  3. 3

    A ConvLSTM layer with kernel size 1  × 1 and number of filters fc1x1, followed by a ConvLSTM layer with kernel size 5  × 5 and number of filters fc5x5,, Hc(X)

  4. 4

    A Maxpooling layer with kernel size 3  × 3, stride of 1 and “same” padding, followed by a ConvLSTM layer with kernel size 1  × 1 and number of filters fd1x1,Hd(X)

These four branches produce intermediate outputs Ha(X),Hb(X),Hc(X), and Hd(X), considering that all the branches receive the same input, X.

These intermediate outputs are then concatenated to form the output to this part, G(X). The second part, residual part, adds both the original input, X, with the output from inception part, G(X), to produce the final output, Y.

The design in Fig. 4 can be described using (6) and (7). “\(\bigodot \)” is concatenation operator. “ +” is addition operator.

$$ G(X)=H_{a}(X) \bigodot H_{b}(X)\bigodot H_{c}(X) \bigodot H_{d}(X) $$
(6)
$$ Y=G(X)+X $$
(7)

Solving HAR using ResIncConvLSTM layer

Figure 5 shows the proposed architecture. The architecture uses ResIncConvLSTM layer as its fundamental module. The architecture is arranged as follows:

  1. 1

    A conventional ConvLSTM layer was used as the lower layer to capture low level features, it also transforms the input shape tensor dimensions from (20, 40, 40, 1) to (20, 40, 40, 24), this dimension expansion improves the process of extracting and learning of new features in subsequent layers, and this expansion cannot be done using ResIncConvLSTM layer as in residual layers the input and the output tensor dimensions should conform,

    Fig. 5
    figure 5

    Proposed model using ResIncConvLSTM layer

    Full size image
  2. 2

    Followed by ResIncConvLSTM, to capture higher level features, then

  3. 3

    Another conventional ConvLSTM layer, this layer reduces the dimensions from (20, 40, 40, 24) to (40, 40, 1); in other words, it returns a single reduced tensor for every input sequence, and finally

  4. 4

    A stack of dense fully connected layers with drop out.

This design is trained against ConvLSTM baseline architecture which is a stack of conventional ConvLSTM layers followed by a stack of dense fully connected layers with dropout proposed in [4]. ConvLSTM baseline architecture is adequate in comparisons because it is simple to make the evaluation of our proposed work unbiased and not affected by any assisting factors.

Both models are trained for 20 epochs using the same dataset. Our model has 52,108 trainable parameters. ConvLSTM baseline model has 70,492 trainable parameters. Tables 1 and 2 show the detailed description of the proposed architecture and ConvLSTM baseline architecture, respectively. The tables describe each layer with its input, input tensor size, and output tensor size. The tables show also whether a ConvLSTM layer returns the whole sequence or not.

Table 1 Proposed model detailed description
Full size table
Table 2 ConvLSTM baseline architecture detailed description
Full size table

Data description

The input is arranged in a 5D tensor form: (B,T,W,H,ch), where B is the batch size, T is the number of frames per sequence, in other words the number of timesteps, W is the width of the frame, H is the height of the frame, and ch is the number of channels. The output is either a 5D tensor with dimensions (B,T,W,H,f) or a 4D tensor with dimensions (B,W,H,f), where f is the number of filters and B, T, W, and H as illustrated before. The 5D tensor case is when there is an output for each timestep. The 4D tensor case is when a single output is returned for the whole sequence. The dimensions of the output depends on the nature of the application. In case of HAR, only one final output, single label, is required for the whole sequence, so the output is either a 4D tensor, or a 5D tensor reduced to a 4D tensor in a subsequent layer. In our proposed model, we chose the latter choice. In other applications, like next frame prediction or object tracking, the output is a 5D tensor because an output is required for each time step.

Results and discussion

This section discusses the setup and the results of the experiments held to show the significance of the proposed work. The experiments are run on a computer with quad-core Intel Core i7 processor running at 2.3 GHz using RAM of 8GB of 1600MHz DDR3 memory.

Datasets

The benchmark used to train our model is KTH dataset [10]. The KTH dataset is a human action video database of six human actions (walking, jogging, running, boxing, two hands waving, and hand clapping). We used KTH in the training and evaluation process because it is one of the biggest human activity dataset. The dataset contains 2391 sequences of around 4 s on average, filmed with a static camera over a homogeneous background, indoor and outdoor settings. The model is tested against both KTH and Weizmann [11] datasets. Weizmann is a human action dataset of 90 sequences of nine actors with ten human actions (walking, running, jumping, galloping sideways, bending, one-hand waving, two-hands waving, jumping in place, jumping jack, and skipping). Only actions similar to the KTH dataset are considered in the evaluation process (walking, running, and two-hands waving). Also, 10% of the KTH dataset is set aside and never used during training to be used in the evaluation process. Also, we trained and tested our approach against a subset of UCF Sports Action dataset. UCF Sports Action dataset is a public dataset collecting a set of 10 actions from different sports that are originally broadcasted on television. The actions included in the dataset are diving, golf swing, kicking, lifting, riding horse, running, skateboarding, swing-bench, swing-side, and walking. We only considered the following actions in the evaluation process: diving, lifting, riding horse, swing-side, and walking. The dataset contains 150 sequence of duration from around 2 to 14 s. Table 3 concludes KTH, Weizmann, and UCF Sports Action datasets description. The training data are sampled at 20 frames per video, resized to grayscale 40  × 40 pixels. The training data is preprocessed, resized, and converted to grayscale frames, to reduce the number of trainable parameters due to our hardware computational limitations. Data augmentation methods are used to expand the dataset. The following data augmentation methods are used: addition of Gaussian noise, flip, shifting, zooming, and rotation by angles between 2 and 12 ∘. Our model is trained and tested on single individual action detection. The model accepts a sequence of frames and outputs a single output for the whole sequence labeling the human activity recognized from the sequence.

Table 3 KTH, Weizmann, and UCF Sports Action datasets’ description
Full size table

Experiments

This section discusses the results of the experiments held on both ResIncConvLSTM-based and ConvLSTM baseline models. Both models are trained for 20 epochs using KTH dataset. Figure 6 shows training and validation losses of our proposed architecture. The figure shows that loss in both training and validation decrease gradually in a similar manner with the successive epochs. This shows that the architecture does not overfit to the training data and actually does learn distinctive features.

Fig. 6
figure 6

Training and validation losses

Full size image

Validation accuracy can be shown in Fig. 7. The figure shows how accuracy of ResIncConvLSTM architecture is lower than ConvLSTM baseline model in the early epochs and then ResIncConvLSTM architecture gradually exceeds ConvLSTM baseline architecture.

Fig. 7
figure 7

Validation accuracy vs. epochs

Full size image

Tables 4 and 5 show the confusion matrices of both ResIncConvLSTM and ConvLSTM baseline architectures on KTH dataset. The tables show that both architectures perform well in recognizing boxing, clapping, and waving activities with accuracies 99.9%, 99.5%, and 99.7%, respectively, for ResIncConvLSTM architecture, and 98.1%, 98.4%, and 98.9%, respectively, for ConvLSTM baseline architecture. ResIncConvLSTM performs slightly better in recognizing boxing, clapping, and waving. Performing well for both architectures is quite expected because each of these activities has distinctive features. ResIncConvLSTM performs better by 10.4%, 14.7%, and 12% in recognizing jogging, running, and walking activities, respectively, with accuracies 79.9%, 82.9%, and 98.1% versus 69.5%, 68.2%, and 86.1%, respectively. The confusion matrix shows how ResIncConvLSTM performance decreases, compared to boxing, clapping, and waving, with activities jogging and running. Both activities are confused with each other and with walking because those activities have similar features. Overall, ResIncConvLSTM performs better than ConvLSTM baseline architecture by 7% with classification accuracies 94.08% versus 87% when tested against KTH dataset. ResIncConvLSTM performs better than ConvLSTM baseline architecture by 11% with classification accuracies 79% versus 68% when tested against Weizmann dataset for the common action classes, walking, running, and two-hands waving. Also, ResIncConvLSTM-based architecture is found to outperform ConvLSTM baseline architecture by 21% when tested against a subset of UCF Sports Action dataset with accuracies 68.8% versus 47.66%. The experiment objective is to prove that the proposed layer, ResIncConvLSTM, still outperforms conventional ConvLSTM. Although, our ResIncConvLSTM-based architecture performed better than ConvLSTM-based architecture in classifying the selected actions, this experiment provide us with the following insight for future research: bigger datasets require deeper networks and bigger input size.

Table 4 ResIncConvLSTM architecture confusion matrix on KTH dataset
Full size table
Table 5 ConvLSTM baseline architecture confusion matrix on KTH dataset
Full size table

Comparison with existing approaches

Table 6 shows a comparison between different state-of-the-art approaches used to solve HAR problem, including our proposed approach, evaluated on KTH dataset. The comparison is held in terms of classification accuracy. Based on the presented comparison, our proposed approach outperforms [19, 20, 29, 30, 34–39, 41, 42] and shows a comparable classification accuracy with [31, 32]. Although classification accuracy of our proposed approach is less than [31, 32], by less than 1%, our proposed approach yields good results using a relatively small number of parameters.

Table 6 Comparison with existing approaches on KTH dataset
Full size table

Conclusions

In this paper, a novel layer based on residual and inception is introduced, ResIncConvLSTM layer. The proposed layer is used for solving HAR problem. A ResIncConvLSTM-based architeture is designed and trained on KTH dataset. The designed architecture is tested on KTH, Weizmann, and UCF Sports Action datasets. ResIncConvLSTM-based architecture is found to perform better than ConvLSTM baseline architecture by 7%, 11%, and 21% on KTH, Weizmann, and UCF Sports Action datasets, respectively. Also, experimental results show the effectiveness of our proposed architecture compared to other state-of-the-art architectures. Our future work will concentrate on applying ResIncConvLSTM on different computer vision applications like segmentation, next frame prediction, object tracking, and scene summarization. We may also investigate dual model approach to solve multiple individual HAR with ResIncConvLSTM.