Human Activity Recognition Based on Motion Sensor Using U-Net

May combine two claims.

On the OPP Gesture dataset, U-Net accuracy is up to 94.7%, 3.3% higher than the second highest score.

“for the OPP Gesture dataset with a large number of short-term activities, the accuracy of the U-Net method is up to 94.7%, 3.3% higher than the second highest score”

May combine two claims.

The Sanitation dataset contains 266555 samples, each with X, Y, and Z three-axis acceleration values.

“The whole dataset contains 266555 samples, in which each sample contains X, Y, and Z three-axis acceleration values.”

May combine two claims.

The Sanitation dataset was collected from an open environment using a triaxial accelerometer worn in a wrist smartwatch at a sampling frequency of 25 Hz, and contains seven types of daily work activity data of sanitation workers.

“The self-collected Sanitation dataset is collected from the open environment.A triaxial accelerometer worn in a wrist smartwatch is used to collect seven types of daily work activity data of sanitation workers.The sampling frequency is 25 Hz.”

May combine two claims.

The U-Net method shows better robustness and excellent performance of recognition on short-term activities and minority classes.

“Moreover, it also shows the better robustness and excellent performance of recognition on the short-term activities and minority classes.”

• Also stated in: § I. INTRODUCTION · plain

  “showing better robustness and recognition performance on the short-term activities and minority classes.”

Probably not a substantial claim: may not pass S3 (load-bearing).

SVM sometimes performs better than the deep learning methods based on sliding window prediction.

“Remarkably, SVM sometimes performs better than the deep learning methods based on the sliding window prediction.”

May combine two claims.

Run and Walk are minority classes of the Sanitation dataset, accounting for 1.5% and 3.5% respectively.

“Considering that Run and Walk are both the minority classes of Sanitation dataset shown in Fig. 5, which account for 1.5% and 3.5% respectively.”

May combine two claims.

In the self-collected Sanitation dataset, there are a large number of substitution and fragmentation errors, which leads to a sharp decrease in the accuracy of FCN and Mask R-CNN.

“In particular, there are a large number of substitution and fragmentation errors in the self-collected Sanitation dataset, which leads to a sharp decrease in the accuracy of FCN and Mask R-CNN.”

Probably not a substantial claim: may not pass S3 (load-bearing).

SVM's performance is even better than that of CNN and CovLSTM on the HAPT dataset.

“For example, the performance of SVM is even better than that of CNN and CovLSTM on HAPT dataset”

May combine two claims.

Although FCN, SegNet, and Mask R-CNN can realize dense prediction, U-Net's performance is still better than theirs.

“It can be found that although FCN, SegNet, and Mask R-CNN can also realize dense prediction, U-Net's performance is still better than that of these comparative methods.”

May combine two claims.

On the OPP Gesture data, LSTM's Underfill and Overfill error rates are 5.54% and 2.74%, while U-Net has lower error rates of 1.42% and 1.39%, respectively.

“Especially for the OPP Gesture data, the Underfill and Overfill error rates of LSTM are 5.54% and 2.74%, while U-Net has the lower error rates of Underfill and Overfill, which account for 1.42% and 1.39%, respectively.”

May combine two claims.

On the Sanitation dataset, U-Net's F1-Score is 0.998 for Run and 0.905 for Walk; the second highest F1-Score for Run is 0.906 achieved by SVM, and that for Walk is 0.78 achieved by CNN.

“Most notably, as for Sanitation dataset, our F1-Score is 0.998 and 0.905 for Run and Walk, in contrasts, the second highest F1-Score for Run is 0.906 achieved by SVM, and that for Walk is 0.78 achieved by CNN.”

May combine two claims.

For the four datasets, U-Net has higher accuracy and shows a significant reduction on Underfill, Overfill, Substitution, and Fragmentation errors compared with the deep learning methods based on sliding window prediction, such as LSTM, CNN, and CovLSTM.

“For the four datasets, U-Net has the higher accuracy and shows a significant reduction on Underfill, Overfill, Substitution, and Fragmentation errors, compared with the deep learning methods based on sliding window prediction, such as LSTM, CNN, and CovLSTM.”

May combine two claims.

The proposed post-correction method improves accuracy on each dataset and significantly reduces the Fragmentation error rate for U-Net, FCN, SegNet, and Mask R-CNN.

“Compared with the original dense prediction results of the four methods, U-Net_PC, FCN_PC, SegNet_PC, and MaskRCNN_PC have improved the accuracy on each dataset and reduced the Fragmentation error rate significantly, which proves the effectiveness of the proposed post-correction method for the dense prediction-based HAR methods.”

May combine two claims.

For WISDM, the recognition Fw-Score firstly improves with the increase of subsequence length and achieves the highest when the subsequence length grows to 244 sampling points, then it decreases.

“For WISDM, the recognition Fw-Score firstly improves with the increase of the subsequence length and achieves the highest when the subsequence length grows to 244 sampling points, then it decreases.”

May combine two claims.

On WISDM, U-Net with block number 5 achieves the highest Fw-Score and has a tendency to remain stable.

“Especially on WISDM, U-Net with the block number of 5 achieves the highest Fw-Score and has a tendency to remain stable.”

May combine two claims.

On the OPP Gesture dataset, the accuracy of U-Net is 94.75%, while the accuracy of U-Net_PC is 94.78%.

“As for the OPP Gesture, the accuracy of U-Net is 94.75%, while the accuracy of U-Net_PC is 94.78%.”

May combine two claims.

U-Net's prediction time is less than 1 second, which can be further improved on better performance machines.

“In addition, U-Net's prediction time is less than 1 second, which can be further improved on better performance machines.”

May combine two claims.

Although the Fragmentation error rate for U-Net_PC decreases, the Underfill error rate for U-Net_PC still increases.

“Although the Fragmentation error rate for U-Net_PC decreases, the Underfill error rate for U-Net_PC still increases.”

May combine two claims.

U-Net outperforms all comparative methods on the four datasets, with better performance on short-term activity recognition and better robustness.

“The experimental results demonstrate that our U-Net outperforms all the comparative methods on the four datasets with better performance on short-term activity recognition and better robustness.”

May combine two claims.

For HAPT and OPP Gesture, U-Net with block number 5 also achieves the highest Fw-Score, but the overall fluctuation range is small with the increase of block number.

“For HAPT and OPP Gesture, U-Net with the block number of 5 also achieves the highest Fw-Score, but the overall fluctuation range is small with the increase of block number.”

Classified as probably not a substantial claim. May combine two claims.

The study used simple cross validation on the four datasets, randomly dividing the whole dataset into 70% training and 30% testing, with a validation set split from the training set at ratio 0.3.

“We adopt the simple cross validation on the four datasets.The whole dataset is randomly divided into the training set and the testing set, in which the training set accounts for 70%.We split the validation set from the training set, and the split ratio is 0.3.”

May combine two claims.

The generated subsequences of densely labeled data have output labels with the same length as input data, whereas traditional sliding subwindows have shorter output labels.

“For each training subsequence, the output labels ŷt has the same length of the input data x t .However, for the training sliding subwindows, the length of the output labels is shorter than that of the input data since all samples in a subwindow just correspond to a single label.”

Probably not a substantial claim: may not pass S4 (based).

Multi-channel time series data can be regarded as a single pixel column, multi-channel image.

“And multi-channel time series data can be regarded as a single pixel column, multi-channel image”

• Also stated in: § 1) DENSE PREDICTION AND U-NET · plain

  “So it can be regarded as a single pixel column and multi-channel image which is input into the U-Net network.”

May combine two claims.

The U-Net combines shallow and deep network information to offset information loss caused by down-sampling, showing better robustness and recognition performance on short-term activities and minority classes.

“Different from the comparative methods, the U-Net combines the shallow and deep network information to offset the information loss caused by the down-sampling operation, showing better robustness and recognition performance on the short-term activities and minority classes.”

May combine two claims.

The paper proposes a HAR framework based on U-Net that performs activity labeling and prediction of each sampling point, achieving dense prediction and overcoming the multi-class window problem.

“In this paper, we propose a novel HAR method based on U-Net to overcome the multi-class problem, performing activity labeling and prediction of each sampling point.”

• Also stated in: § I. INTRODUCTION · plain

  “We propose a novel HAR framework based on U-Net to overcome the multi-class problem.”

Probably not a substantial claim: may not pass S4 (based).

U-Net network is applied to HAR for the first time.

“To the best of our knowledge, U-Net network is applied to HAR for the first time”

May combine two claims.

Dense labeling, labeling every sampling point instead of sliding window labeling, preserves correct label information for each sample and improves recognition accuracy of the classifier.

“Therefore, we adopt dense labeling by labeling every sampling point instead of sliding window labeling so as to preserve the correct label information for each sample and improve the recognition accuracy of the classifier.”

May combine two claims.

Existing HAR models based on CNN architecture cannot achieve dense prediction for each sampling point's label in time series data due to max pooling reducing top-level output resolution and causing size mismatch; they are only suitable for sliding window prediction.

“The maximum pooling operations in the traditional CNN architecture result in reducing the resolution of toplevel output and a size mismatch between the output and the input.Due to the size mismatch, the existing HAR model based on the CNN architecture cannot achieve dense prediction for each sampling point's label in the time series data, only suitable for sliding window prediction.”

May combine two claims.

Because samples within a sliding window may not share the same label, sliding window labeling can mislabel samples and reduce recognition accuracy.

“Considering that all samples of a window may not always share the same label, some samples will be mislabeled.Both the two methods may lead to incorrect labeling, thereby reducing the recognition accuracy.”

• cites [13] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The activity-based misalignment measure in prior work focuses on classification of interesting classes and the NULL class, with all misalignment types related to wrong classification of the NULL class.

“To analyze and evaluate the continuous activity recognition on time series data, the activity-based misalignment measure is proposed in [32], where the HAR model focuses on the classification of the interesting classes and the NULL class, all the misalignment types are related to the wrong classification of the NULL class.”

• cites [32] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The U-Net network settings (subsequence length 224 and network block number 5) were selected based on validation procedures on the WISDM dataset and are consistent across all datasets.

“As for the parameter tuning of the network block number and the subsequence length, we select the optimal parameters based on the validation procedures on WISDM dataset.The subsequence length is set to 224, the network block number is set to 5. Considering the good robustness of the deep network, the network settings of our U-Net are consistent across all datasets.”

• AssumptionRests on an assumption: Deep networks have good robustness such that the U-Net network settings can be consistent across all datasets.

Classified as probably not a substantial claim. May combine two claims.

The optimal hyper-parameters of each deep network were determined by validation procedures on the WISDM dataset, and the settings of each deep network were consistent across the four datasets.

“The optimal hyper-parameters of each deep network are determined by the validation procedures on WISDM dataset.The setting of each deep network is consistent in the four datasets.”

• AssumptionRests on an assumption: Optimal hyper-parameters selected on the WISDM dataset generalize to the other three datasets.

May combine two claims.

The multi-class window problem is common in motion sensor HAR and significantly negatively impacts recognition of short-term activity sequences, making short-term activity classification challenging.

“This problem caused by fixed sliding window segmentation and labeling is called multi-class window problem [14].Multi-class window problem is a common problem in HAR based on the motion sensor, which has a significantly negative impact on the recognition of short-term activity sequences.It makes short-term activity classification challenging.”

• cites [14]Not checked yet

  Not checked in the free run.

May combine two claims.

Current machine learning and deep learning HAR methods use fixed sliding window segmentation and label all samples in each window.

“Traditional human activity recognition (HAR) based on a motion sensor adopts sliding window labeling and prediction.”

• Also stated in: § I. INTRODUCTION · plain

  “At present, both the machine learning-based and the deep learning-based HAR methods use fixed sliding window technology to divide the sensor signal into fixedlength windows and label all samples in the window with the labeling strategies.”

• cites [12]Not checked yet

  Not checked in the free run.

May combine two claims.

The post-correction algorithm focuses on correcting Fragmentation and Substitution errors because Overfill and Underfill errors are difficult to observe without groundtruth.

“Considering that the errors of Overfill and Underfill are difficult to be observed in the dense prediction results without knowing the groundtruth, our algorithm focuses on correcting the two types of errors: Fragmentation and Substitution.”

• AssumptionRests on an assumption: The boundary and adjacent prediction probability determine the correct activity class for correcting Substitution errors.

May combine two claims.

Theoretically, the dense prediction results of FCN and Mask R-CNN are rough and not as good as U-Net and even worse than traditional CNN based on sliding window prediction.

“Theoretically, the dense prediction results of FCN and Mask R-CNN are rough and not as good as U-Net and even worse than traditional CNN based on sliding window prediction.”

May combine two claims.

U-Net is suitable for the unbalanced dataset with reliable performance on the minority classes.

“This also proves that our U-Net method is also suitable for the unbalanced dataset with reliable performance on the minority classes.”

May combine two claims.

SVM's performance is not stable enough and has inferior robustness.

“but the performance of SVM is not stable enough with inferior robustness.”

May combine two claims.

U-NET can detect more details of the transition activities and has more stable performance.

“This shows that U-NET can detect more details of the transition activities and has more stable performance.”

The quote only partly states this claim.

Dense prediction has an advantage over sliding window prediction.

“which further proves the advantage of dense prediction.”

May combine two claims.

LSTM and CovLSTM adopt the sliding window method, which brings the multi-class window problem and reduces accuracy, especially in short-term activity recognition.

“The reason is that LSTM and CovLSTM adopt the sliding window method, which brings the multi-class window problem and reduces accuracy, especially in short-term activity recognition.”

May combine two claims.

U-Net is based on dense prediction and avoids the multi-class problem, thus it improves recognition accuracy and is more suitable for short-term activity recognition.

“Our U-Net method is based on dense prediction and avoids the multi-class problem, thus it further improves the recognition accuracy and is also more suitable for short-term activity recognition.”

May combine two claims.

U-Net is suitable for simple long-term activity recognition and more suitable for short-term activity recognition, showing strong robustness.

“This proves that U-Net method is not only suitable for simple long-term activity recognition, but also more suitable for short-term activity recognition, showing strong robustness.”

May combine two claims.

U-Net utilizes high-resolution feature map information from the encoder instead of reusing pooling indices only, indicating that the entire features from the encoder contain fine-grained information important for dense prediction.

“Unlike the SegNet method for dense prediction, our U-Net utilizes the high-resolution feature map information from the encoder instead reusing pooling indices only, this indicates that the entire features from the encoder contain finegrained information which plays an important role in dense prediction.”

May combine two claims.

FCN's performance is not stable and still has the potential for significant improvement.

“Though FCN outperforms our U-Net method in activity Sit_to_Stand and Lie_to_Sit, the performance is not stable and still has the potential for significant improvement.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The comparative methods adopt the same data preprocessing procedure as the U-Net method, and the subsequence length is consistent with U-Net's setting, except for Mask R-CNN.

“The comparative methods adopt the same data preprocessing procedure as our U-Net method, the subsequence length also keeps consistent with the U-Net's setting, except the Mask R-CNN.”

May combine two claims.

U-Net results are not greatly affected by the length of the subsequence, showing the effectiveness of the design and great robustness.

“U-Net with the subsequence length of 224 also achieves the highest Fw-Score on the other datasets and the results are not greatly affected by the length of the subsequence, showing the effectiveness of our design and great robustness.”

May combine two claims.

U-Net overcomes the multi-class window problem inherent in the sliding window method and realizes the prediction of each sampling point's label in time series data.

“Different from the existing machine learning-based and deep learning-based HAR methods which use the sliding window labeling and prediction, our U-Net method overcomes the multi-class window problem inherent in the sliding window method and realizes the prediction of each sampling point's label in time series data.”