Have you ever considered that a smartwatch or your phone could play a crucial role in exonerating someone from a murder charge? This real-life story unfolded in the case of Nicole VanderHeyden in 2016.
Douglas Detrie, Nicole's husband, was initially identified as the prime suspect. However, in 2018, unexpected developments occurred in the case.
Prosecutors decided to analyze the data from Nicole's Fitbit smartwatch, a device she had gifted to Doug. This smartwatch was capable of recording precise information about physical activities, including:
- Step counts
- Heart rate
- Sleep patterns
The discoveries made from this data led to a dramatic shift in the case. Analysis of the Fitbit data revealed that Doug had not engaged in significant physical activity at the time of the crime and was actually asleep.
For more details about this fascinating case, visit Oxygen's coverage.
This project explores the impact of wearable technology in forensic investigations and legal cases.
In this project, we’re leveraging data from smartphone accelerometers to analyze and identify human physical activities. Accelerometers are sensors found in many smartphones and wearable devices, providing detailed information about movements.
Our goal is to develop a machine learning model that can accurately detect and classify various physical activities based on accelerometer data. We aim for high accuracy and reliability in recognizing activities, including:
- Triaxial acceleration from the accelerometer (total and body acceleration)
- Triaxial angular velocity from the gyroscope
- A 561-feature vector with time and frequency domain variables The activity label for each recorded activity An identifier for the individual who performed the activity
We’ll be using the "Human Activity Recognition with Smartphones" dataset available on Kaggle. This dataset includes:
Triaxial acceleration from the accelerometer and body acceleration estimates Triaxial angular velocity from the gyroscope A 561-feature vector with time and frequency domain variables Activity labels for each recorded activity An identifier for each individual For more information, visit the dataset on Kaggle Data set.
This project is developed with the aim of analyzing and predicting data using machine learning techniques. In this project, two distinct models of popular machine learning algorithms are employed: Random Forest and K-Means
Interpretation of the Graphs:
Each graph illustrates how the gravitational angle in the Y direction (angle(Y, gravityMean)) changes over time for specific activities. These variations help distinguish between different activities:
Laying Down (LAYING): The graph shows relatively constant or oscillatory patterns, indicating the body is in a stable position.
Walking (WALKING): The graph displays periodic peaks and troughs, reflecting bodily movement during walking.
Walking Upstairs (WALKING UPSTAIRS) and Walking Downstairs (WALKING DOWNSTAIRS): These graphs exhibit sudden changes due to the effort involved in ascending or descending stairs.
In this analysis, we employed machine learning models to predict human movements based on sensor data. Specifically, we utilized the Random Forest model for classifying and predicting human gestures. This model processes data by focusing on features with the highest predictive power, enabling effective identification and analysis of movement patterns from the sensor data.
Random Forest offers several advantages over single decision trees: Reduced Sensitivity to New Data: Random Forest is less sensitive to new data, providing more stable and reliable predictions. Enhanced Accuracy: It typically achieves higher accuracy due to its ensemble approach, combining predictions from multiple decision trees. Robustness to Outliers: Random Forest is resilient to outliers and noisy data, which improves overall model performance. Automatic Feature Selection: The model automatically identifies and prioritizes the most influential features for prediction, streamlining the analysis process.
The features displayed in this chart are ranked based on their impact on the decision-making of the Random Forest model. The higher the importance value, the greater the influence of that feature on predicting human activities.
For example, the gravityMean feature is identified as the most significant, indicating the effect of gravitational force on physical behaviors. Accelerometer-related features, such as tBodyAcc-min()-X and tBodyAcc-energy()-X, are among the most important variables, assisting the model in distinguishing between movements like standing and walking. Additionally, gyroscope-related features, such as fBodyAcc-max()-X and fBodyAcc-min()-Y, are also among the top features and help the model in detecting rotational and twisting body movement
In our implementation of the Random Forest model, we used cross-validation to evaluate its performance. The hyperparameters of the model were tuned using the cross-validation results to optimize its performance on the dataset. The hyperparameters are detailed in the provided DataFrame, ensuring that the model is well-tuned for the task at hand.
In the K-Means algorithm, our goal is not to predict but to group data points that share significant similarities. By doing so, we can identify the common features within each group and also determine the characteristics that differentiate the data points across various groups. This approach helps us uncover hidden patterns in the data, enabling us to perform deeper analyses of the data's structure.
One of the key steps in this model is determining the optimal number of clusters. To do this, we run the model with different numbers of clusters and calculate the distance of each data point from the centroid of the nearest cluster. This process allows us to identify the optimal number of clusters, ensuring that the clustering best reflects the structure of the data
By analyzing this chart and using the "Elbow Method," we observe that the rate of inertia reduction slows down after 3 to 5 clusters. This behavior indicates that increasing the number of clusters beyond this range does not significantly improve the quality of clustering. Therefore, the optimal number of clusters lies between 3 and 5. Choosing the number of clusters within this range allows us to segment the data effectively and identify patterns with greater accuracy
In this test, our goal is to evaluate the quality of the data clustering. This assessment determines whether each data point has been correctly placed in the appropriate cluster. The Silhouette Score indicates how close the data points are to their own cluster and how far they are from other clusters, providing a reliable measure of clustering quality.
We observe that selecting 2 clusters yields the highest Silhouette score. However, considering our goal of dividing the data into more clusters, choosing 3 clusters seems more reasonable. This choice still maintains an acceptable Silhouette score while allowing for a more detailed clustering of the data.
In this plot, we have visually represented the clusters using PAC, which is employed for cluster analysis. This method helps us to better understand the structure and relationships between clusters and provides a clearer view of the shared features and differences among them.
In this section, we use an evaluation called the confusion matrix to assess the performance of our model. This evaluation helps us understand how the model assigns input data to different clusters and allows us to evaluate its effectiveness in accurately identifying activities.
Cluster 0: Contains a high number of walking-related activities (WALKING and WALKING_UPSTAIRS).
Cluster 1: Focuses on non-walking activities (LAYING, SITTING, STANDING).
Cluster 2: Includes walking activities with a significant number of WALKING_DOWNSTAIRS samples