The dissertation's major motivation was to create a precise and accurate indoor navigation system by manipulating an impulsive received signal strength indicator (RSSI) in a very challenging underground car park environment. To achieve our goals a testbed architecture and fingerprint database were created by recording and pre-processing sensor data by using a combination of solar-powered blue tooth sensors and the BLE-USB bridge receiver. A qualitative analysis was then conducted to evaluate its efficiency. This fingerprint database was further used with prominent machine learning techniques, pathfinding algorithms, floor maps, and an audio module to create a navigation application. Various machine learning algorithms like affinity propagation, k-nearest neighbours, k-means, support-vector machine, and logistic regression were studied in different environments, and affinity propagation was finally used due to its reliability and robust nature, to successfully navigate around the testbed. The entire process is divided into 2 phases: Online Phase and Offline Phase.

Each reference point has 5 measurement points, each resembling a cluster labeled P1, P2, P3, P4, and P5. The fingerprint database will be created by recording sensor data from each measurement point. It will become another crucial dataset label, which will help in determining the position of the user. X and Y coordinates are recorded for each measurement point. Measurements are manually taken with the help of 50-meter open reel tape.

The above python file in the "Code Base/1 Data Preprocessing and Visualization" is used to extract the RSSI value and preprocess the dataset.

The above file in the "Code Base/1 Data Preprocessing and Visualization" is used to combine all individual measurement point specific datset into single Master dataset.

These file are used to further preproces the dataset to make it suitable for affinity propagation, coarse localization and Kmeans models. In this cluster are created and cluster details are stored along with master data in CSV files."Code Base/1 Data Preprocessing and Visualization" contains these files.

These file are used to store the created CSV files into respective databases."Code Base/1 Data Preprocessing and Visualization" contains these files.

This file was initial used to check the useability of the master data containing the cluster informations. "Code Base/1 Data Preprocessing and Visualization" contains these files.

Below Images shows the 3D scatter plot generated during visualization. It is evident from Figures that signal strength of access points 1 and 4 is maximum around their corresponding blocks.

This file present in "Code Base/1 Data Preprocessing and Visualization" was used to visualize the quality of collected master data.

To analyze the effect of the environment on the model prediction, we have created various test scenarios. These test scenarios include test data collected from the testbed at various instances with different vehicle intensity levels. Multiple scenarios have been created as follows,

• Scenario 1: Base scenarios as, using the original dataset by dividing it into train and test dataset.
• Scenario 2: Test data as fingerprints collected when 14 vehicles were present in the dataset.
• Scenario 3: Test data as fingerprints collected when testbed had vehicle intensity as 23.
• Scenario 4: In our fourth scenario, we use a combination of 21 vehicle and original dataset to train models and use the 13 vehicle data to test the results.
  The different scenarios are taken to study various problems in RSSI based navigation.

We use the scenarios mentioned above and datasets to compute the performance of multiple models. We use distance-based models as evaluation models as RSSI localization is based on distance. Models such as K-Nearest Neighbour, K-Means, Stare Vector Machine, Logistic Regression, and Affinity propagation were studied and analyzed.

Above source code files present in "Code Base/2 Model Evaluation/" is used to evaluate performance of RSSI based preiction using different Machine learning models. Accuracy, Precision Recall and Prediction Errors are calculated to find the best Model. After carefull analysis we selected Affinity Propagation as our final model. The following figures shows the position error of the final selected model.

The above python files in the "Code Base/3 Visual Module/" were used to create Visual Floormap module.

The above python files in the "Code Base/4 Audio Module/" were used to create Audio guiding module.

Here we will discuss step by step functioning of our user application. The online phase's generated fingerprints are stored in the database as two separate tables, ‘master_data’ and aggregated ‘affinity_prop_master’ with 16,824 and 220 data, respectively. This data acts as a base for the offline application phase. The following step describes the application workflow. *Step 1: The user collects the current RSSI data from the current location using the Cypress BLE beacon app and stores it in a CSV file. *Step 2: The application initially trains the selected model with an aggregated ‘affinity_prop_master’ dataset. This is the Coarse localization stage. *Step 3: The user’s current location is predicted using the current RSSI readings using the model. This becomes the source node. *Step 4: The user is prompted to enter the destination location. This is considered as the destination node. *Step 5: Dijkstra’s pathfinding algorithm is used to compute the shortest path and calculate the distance from the source to the destination node. *Step 6: The selected model is then trained with the cluster member present in the ‘master_data’ dataset for the current Reference Point. *Step 7: The user’s actual location is predicted, and the geo-coordinate for the user is computed. This constitutes the Fine localization stage. *Step 8: Audio-Visual module is then used to locate the user, find the destination, and guide the user to the desired location. *Step 9: The user traverses forward towards the prompted location and tries to collect the new RSSI readings. *Step 10: If the current location matches the destination location, then “Destination Reached” else “Go to Step: 3”. Thus, by integrating all the individual modules, the application is designed to allow users to navigate the indoor environment freely.

Finally we have successfully navigated using 4 differnt models. Thought Affinity Propagation was superior than other models, Hence chosen as our final model. The above files have the final navigation apllication codes. "Code Base/5 Navigation Models/" contains these files.