From: Detecting activity locations from raw GPS data: a novel kernel-based algorithm

Criteria |
A_{ft}
| Comment |
A_{kd}
| Comment |
---|---|---|---|---|

Highest proportion of tracks with correctly identified number of stops. depending on parameter value | 65.5% | Obtained with 1000 m radius | 92.3% | Obtained with 200 m bandwidth |

Number of noise/parameter combinations for which detection correctly identifies three stops for at least 70% of tracks (out of 24 combinations) | 3 | Performance sharply decreasing with increasing noise; best combination yields 75.6% of correct identification of three-stop tracks | 15 | 10 out of these 15 successfull combinations with correct detection of 90% or more of three-stop tracks |

Number of correctly identified stops among tracks with close (<800 m) neighbours | 132 | Larger radii=better prediction | 194 | Inversed U-shaped relation to bandwidth: best capacity with ‘average’ bandwidth of 200 m |

Number of noise/parameter combinations for which the average number of detected stops is around 3 (2.8<average<3.2) | 6 | 10 noise/parameter combinations for which average=zero | 15 | 2 noise/parameter combinations for which average=zero |

Number of noise/parameter combinations for which distance between detected and true stop is less than 15 m in average (out of 24 combinations) | 8 | Standard-errors larger in AFT than in AKD for all combinations | 17 | 11 combinations with less than 10 m in average |

Number of noise/parameter combinations with duration difference between detected and true stop less than 10% error | 11 | AKD outperforms AFT for 16 out of 24 combinations | 16 | Duration difference below 5% for 200 m bandwidth at all noise levels |