3.1. Experimental Results
In Table 1
, the time needed to detect the casualty in each trial is shown, about one hour on average. We successfully detected the casualty in eight out of nine trials performed. Being fast and precise in casualty detection is a key factor because 80% of survivors are recovered alive if rescued within 48 h.
The results of each trial are shown Figure 4
, Figure 5
and Figure 6
and described and commented on in the rest of this section. O2
is measured as concentration, while CO2
is in parts-per-million (ppm). Because the CO2
sensor is not calibrated, the CO2
data do not represent the real concentration and the absolute measured values in each trial vary widely depending on the time the measurement was taken and the environment around the site. For this reason, relative variations of CO2
during trials were considered, and further confirmation from a rescuer or other sensors was required to verify the presence of the casualty in that specific area. The areas with relatively high levels of CO2
are indicated in yellow. Areas manually checked with a thermal camera are circled in purple.
shows the results of the first day’s trials.
Day 1, morning trial: The gas sensor located several possible locations for the casualty. The reason for those abnormal concentrations is that the person reached the center of the site through tunnels in the test sites (C5, A5, A8, A10, and A11 are sections of the same tunnel). The thermal camera images confirmed the presence of the casualty in the estimated area indicated by the red square, in the square composed by cells B9, C9, B10, and C10.
Day 1, afternoon trial: The gas sensor identified an area with a peak CO2 concentration and the thermal camera confirmed the presence of the casualty in the area indicated by the gas sensor data, in the square composed by cells B7, C7, B8, and C8.
Day 1, evening trial: In this test, the casualty was located in the square composed by cells B11, C11, B12, and C12, using only the thermal camera. The gas sensor did not work properly because the affected area is a large area in which the wind could easily change the CO2 concentrations, so the presence of a casualty did not significantly change the CO2 concentration in this situation. This test was useful to analyze the factors that can lead to localization failures when using a gas sensor. However, this kind of area can be easily searched by a rescue team or a rescue dog because it is near the boundaries of the disaster area, outside the collapsed structure.
shows the results of the second day trials.
Day 2, morning: Both the gas sensor and the thermal camera located the casualty. The C11, D12, C12, D12 area is part of a corner in which the gas concentration was unusually high, and the camera could be inserted through a hole in the rubble to verify the presence of the casualty.
Day 2, afternoon: Both the gas sensor and the thermal camera located the casualty in the square composed by cells C6, D6, C7, and D7 that was beside a wall in a corridor where the gas sensors and the camera could be placed. It is important to note that, in this case, the gas sensor detected a high concentration of CO2 in the whole corridor, so the exact position of the casualty could only be confirmed with a thermal camera.
Day 2, evening: This was the only trial in which the sensor system failed to locate the casualty. A high concentration of CO2 was found in the area around B2, C2, B3, and C3, but the presence of many obstacles obstructing the view made verification via thermal camera impossible. This area is a maze of corridors in a semi-closed area with low air circulation, with the possible presence of grass and animals that might raise the concentration of CO2. Moreover, the corridor in C2 was not reachable by gas sensors on the telescopic pole.
shows the results of the last day’s trials.
Day 3, morning: The gas sensor found a high CO2 concentration very close to the casualty. However, the presence of many obstacles obstructing the view made verification via thermal camera impossible, so the casualty was located in the square composed by cells B3, C3, B4, and C4 based only on the gas sensor data.
Day 3, afternoon: The casualty was located very fast because the gas sensor measured a relatively high level of CO2 in the square composed by cells B11, C11, B12, and C12 and the thermal camera confirmed the presence of the casualty through a hole in the corridor.
Day 3, evening: The casualty was located in the square composed by cells B9, C9, B10, and C10 using only the thermal camera. The data from the gas sensor were corrupted because of hardware problems on the gas sensor board.
3.3. Evaluation of Microphone and Audio Processing Algorithm
In all the experimental trials, the casualty was supposedly unconscious. Therefore, the person did not speak or produce other sounds such as scratching during the whole trial. However, in real disaster scenarios, there are cases in which the casualty is not unconscious and can produce sounds. For this reason, an algorithm for the detection of sounds that might be related to the presence of a casualty was designed and tested. The hardest problem was to make the algorithm less sensitive to background noise. A disaster site is often a noisy environment, with people searching for victims, vehicles, and various natural and artificial sounds. A dynamic threshold for the classification between a possible sign of life and background noise, based on the average level of sound in the area, was proposed. Of course, this method implies that in extremely noisy environments the detection of feeble sounds will not be possible. However, in this way the system is more robust and automatically rejects sounds that are not linked with the presence of casualties, reducing the number of sounds that must be listened for to check the presence of casualty in a specific area. In particular, speech has characteristic features that were used to separate it from other suspect noises. Figure 8
shows the results of the Day 3 afternoon test, in which we spoke directly to the casualty after locating them to test the audio recognition system. The microphone was placed on a telescopic pole and inserted in a hole in the same corridor where the person was detected by using the gas sensor and the camera. Then, we asked the casualty to perform three different tests: to not move and stay in silence while we talked outside, to call for help at a low volume inaudible by the human ear from outside, and to simply scratch on the ground. Audio detection results are shown in Figure 8
. Moreover, we detected an unwanted cough, confirming the presence of the casualty, in the area with a high level of CO2
during the Day 1 afternoon trial, and another suspect noise during the Day 2 afternoon trial, when the person moved into the corridor.
The result of audio detection performance evaluation is shown in Table 3
. The proposed algorithm can automatically differentiate the sound data and save it in different folders. In Table 3
, the first row α/β represents the correct sound identification rate, where β is the total number of automatically classified sound files present in each category folder and α is the correctly classified number of sound files, which were validated manually.
The correct voice recognition rate is 89.36% in a noisy environment. The correct classification rate for human-related suspect noise, including scratching and coughing, is 93.85%. Therefore, using a microphone in connection with other sensors would be beneficial for the detection of casualties.