Recognition of vehicle models from engine sounds

In this thesis study, it is aimed to recognize the vehicle models of 5 different model vehicles using their own motor sounds by using sound processing and classification techniques. Sound recordings were not taken in an isolated and enclosed environment. For this reason, the sound recordings were taken in an open environment at night, where the ambient sounds (human voice, bird sound, ambient sounds, etc.) are expected to be minimal. All the vehicles were operated in idle mode while sound recordings were taken. While the vehicles were operating in idle mode, 50 voice sounds were collected every 10 seconds for each car. The collected sound data was digitized by being transfered the computer. After examining the sound data, the power spectral densities of the signals were calculated by applying the welch method to 50 sound data of each interval and the graphs were ploted separately for all sounds. 17 different frequency regions were determined from the power spectrum graphs. The amplitude values in this frequency domain are taken as attributes for each vehicle. When there are 50 voice recordings and 17 feature vectors for each vehicle, the feature matrix is set to 17x50. Since there were 5 different vehicles in the study, a total of 17x250 feature matrices were obtained. The 17x125 part of this matrix was used in training for classification and the remaining 17x125 was used to calculate success in the trained classification. Artificial Neural Networks, Support Vector Machines and k-Nearest Neighbors method were used for classification. 100% success rate was achieved by Support Vector Machine method and 99.2% success rate was achieved by Artificial Neural Networks and k- Nearest Neighbors methods. With these high success rates, the vehicle models were recognized.