Hybrid and Other Techniques

Hybrid and Other Techniques

Meshlium

Meshlium is a device that uses WiFi and Bluetooth scanners to detect other devices, which can be used for a range of applications/research (e.g. Vehicle Traffic Detection).

WiFi and Bluetooth radios (of devices) periodically send out messages, containing:

• MAC address of wireless interface
• Strength of the signal (RSSI)
• Vendor of the smartphone
• WiFi Access Point and Bluetooth friendly name
• Class of Device (CoD) (only when Bluetooth)

Figure 1: Meshlium Summary

MAC address randomization: for privacy randomized MAC address, reverts to "factory" MAC address when connected to WiFi

Adaptive Frequency Hopping (AFH): algorithm that enables Bluetooth radio to dynamically identify channels already in use and avoid them

Visual Based Indoor Localisation

• Image processing module: interpret visual data
  • Object recognition in images also help to contextualise the location
    • Eg: stove, fridge and sink indicate the place is likely a kitchen
  • Local feature descriptors: detect local interest points in an image, describe and store them as words
    • Bag-of-words model then compares the collection of word descriptors with the map (trained with places also described by bag-of-words) to match a place
    • Local features do not take the overall geometry into account
      • Thus pose-invariant: place is recognizable regardless of the position/orientation of the source image
      • However, adding geometric information improves robustness of place matching
  • Global feature descriptors: create a fingerprint of a location based on detected features
    • Uses color histograms, feature detection (edges, corners, color patches)
    • These features are ordered from \(0^o\) to \(360^o\) into a fingerprint, using omni-directional cameras during training phase
    • Assumes the input live data is at similar height/location of the training data
  • Generally, combining both local and global descriptors provides the best results
• Map: maintains a representation of knowledge of the world
  • Usually a relational (topological/cognitive) map rather than absolute/geometric positioning
  • Consists of bounded places
    • Place signature: set of visual information that distinguishes it from other places
    • Gateway: physical boundaries of a place, where the physical appearances changes significantly
  • Methods:
    • Pure image retrieval: matching based only on image similarities, no position information is required/given
    • Pure topological map: stores relative positions of places, no metric information stored
      • Speeds up searching as indexing is possible
    • Topological-metric / Topometric map: enhance topological maps with direction and/or distance
      • Appearance-based option: metric information only between places, not within places
      • Sparse landmark option: metric information extracted from depth values between key landmarks inside the image
      • Dense occupancy grid option: same as sparse landmark but for more feature points, more GPU/memory-intensive
• Belief generation: combines information from above components to make decision on place familiarity
  • Bag-of-words model: TF-IDF scoring (term frequency - inverse document frequency)
    • Each visual word in image is scored by frequency of it appearing in image, against how common the word is across all images
  • Voting scheme: use multiple data streams to vote confidence of matching
    • Eg: multiple color bands that give unanimous voting and confidence value > threshold
  • Artifical neural network: Continuous attractor network (CAN)
    • Mimic neural network of a rat hippocampus using local excitation and global inhibition layers

Changing Environments

• Image processing module:
  • Invariant methods: focus on features that are invariant despite changing environments
    • Eg: edges and corners remain prominent despite lighting changes
    • This is also true for convolutional neural networks: mid-level features are robust to changes in the environment
    • Alternatively, use training images that are as 'change-invariant' as possible, or pre-process live data to reduce changes
  • Learning methods: define the relationship between how a place can appear across different times
    • Eg: use pairs of images between two different seasons, or day and night
• Map: how to deal with different representations of the same place
  • Remember and forget: Balance between new observations (that may be fleeting/inconsequential) and overwriting obsolete information
  • Multiple representations of the same environment: to capture cyclic/regular changes in environment
    • Eg: seasonal changes are cyclic
    • Store information of the same place (or whole-map level) at different instances of the required timescale