TL;DR
This paper introduces a "thermalfootprint" model that uses thermal camera data to estimate when someone left a location by analyzing the cooling rate of their footprints. The model, based on Newton's Law of Cooling, shows promising accuracy for short time windows after departure, offering a novel tool for forensic investigations.
What Problem Does This Paper Address?
In criminal investigations, determining the precise time of departure from a crime scene can be crucial for narrowing down suspects and reconstructing events. Traditional methods may be limited by the complexity of crime scenes or the need for direct observation. This paper addresses the challenge of extracting temporal information from subtle, often invisible, heat traces left behind by individuals. Specifically, it focuses on the heat signature left on the ground by footprints, which can be captured by thermal cameras but are imperceptible to the naked eye. The goal is to develop a reliable method to estimate departure time from these heat traces, especially in cases where direct evidence is scarce.
Methods
The core of this research is the development of a "thermalfootprint" model. The process begins with acquiring thermal videos of the heat traces left on the floor. These videos are then processed to extract one-dimensional thermal footprint signals. The key insight is that the residual heat from a footprint on the ground follows an exponential decay pattern, as described by Newton's Law of Cooling.
The authors developed algorithms to:
- Identify the "departure point": This is the point where the thermal signature begins to significantly change, indicating the moment the person left.
- Determine the "starting point of exponential decline": This marks the beginning of the predictable cooling phase.
- Establish the "thermalfootprint model": This model mathematically describes the exponential decrease in temperature over time, using parameters derived from the observed heat decay. They employed a combination of statistical analysis and optimization techniques (Maximum Likelihood and Least Squares) to fit this model to the data.
Validation experiments were conducted by comparing the model's estimated departure times against ground-truth times (obtained via stopwatch) and through subjective estimations by human observers.
Datasets
The study utilized a dataset of over 100 thermal curve sets collected from experimental participants. These participants stood on a wooden floor for a period of 30 to 180 seconds while wearing socks. The ambient temperature during these experiments ranged from 16°C to 28°C. No new benchmark datasets were introduced; instead, existing thermal imaging data was collected and analyzed.
Key Insights and Ideas
- Footprints as Time Capsules: The residual heat signature left by footprints on a surface can act as a temporal marker, decaying predictably over time.
- Newton's Law of Cooling is Key: The exponential decrease in temperature of the footprint aligns well with Newton's Law of Cooling, providing a strong physical basis for the model.
- Short Windows are Best: The model's accuracy is significantly higher when estimating departure times within a short window (e.g., minutes) after the person has left. As time progresses, the heat dissipates, making accurate estimation much harder.
- Subjective vs. Objective: While human estimation can be useful, the thermalfootprint model offers a more objective and potentially more accurate method, especially in the initial stages after departure.
- Factors Affecting Accuracy: The study highlights that factors like the selection of the Region of Interest (ROI), ROI size, and environmental temperature can influence the model's performance.
Key Results
- High Model Fit: The thermalfoot models achieved correlation coefficients generally above 0.99, indicating a good fit to the observed heat decay data.
- Residual Analysis: In validation experiments, residuals between estimated and ground-truth departure times were mostly within ±150 seconds.
- Accuracy Varies with Time: The reverse accuracy of the model for estimating departure time decreased significantly as the capture time point moved further from the departure event:
- One-third capture time: 71.96%
- One-half capture time: 50.47%
- Two-thirds capture time: 42.06%
- Three-fourths capture time: 31.78%
- Four-fifths capture time: 21.70%
- Five-sixths capture time: 11.21%
- Comparison with Subjective Methods: The thermalfoot model generally outperformed subjective estimation methods, especially when more thermal information was available. However, for later capture times, the accuracies of the model and subjective methods became comparable.
- Short-Term Performance: The model demonstrated good performance for departure time reversal within a short time window.
- Long-Term Limitations: Accuracy drops to approximately 15% for estimating departure times in a long time window.
Limitations and Future Work
The authors acknowledge several limitations and suggest future research directions:
- Outlier Influence: The presence of outliers (data points with large deviations) can affect model accuracy. Future work could explore robust outlier detection and handling methods.
- ROI Selection: The choice of Region of Interest (ROI) and its size significantly impacts the generated thermal curves. Further research is needed to optimize ROI selection strategies.
- Environmental Temperature: While the study was conducted within a specific temperature range (16°C to 28°C), the model's performance under different ambient temperatures, especially colder conditions, needs further investigation.
- Failure Point: The study identified a "failure point" beyond which the model's accuracy drops below an acceptable threshold (15%). Understanding and potentially mitigating the impact of this failure point is crucial.
- Practical Application Challenges: In real-world scenarios, determining the exact starting point of the exponential decline period and lag time can be challenging without prior knowledge. The current approach relies on statistical assumptions for these parameters.
- Time Window: The model is most effective within a short time window. Extending its utility to longer timeframes or integrating it with other forensic techniques could be beneficial.
Summary updated