A novel approach for hardware based sound approaching detection
演題番号 : P1-r16
Mauricio Kugler:1 Hirofumi Tsuzuki:1 Susumu Kuroyanagi:1 Akira Iwata:1
1:Department of Computer Science & Engineering, Nagoya Institute of Technology, Nagoya, Japan
The detection of approaching vehicles is a very important topic on the development of complementary traffic safety systems. However, the majority of the proposed approaches are very complex and not suitable for embedded applications. This research proposes a new sound approaching detection algorithm specifically intended for implementation in an FPGA device. This system consists of three blocks: signal processing, feature extraction and classification. On the first, the sound signal is sampled at 48kHz and divided in N=5 frequency channels between 1kHz and 2kHz, corresponding to the spectrum of the tyres' noise in asphalt. If the sound energy variation is positive, the sound source is approaching the subject, otherwise, it is stopped or getting away of it. The energy for each nth frequency channel can be measured by counting the number of spikes xn(t) in a time window of W samples. The differences of all neighbor pairs of S consecutive windows are calculated. If xn(t+sW)-xn(t+(s-1)W) > 0, the feature vector bit Fn,s is set to 1, otherwise 0. If |xn(t+sW)-xn(t+(s-1)W)|>β, then Vn,s is set to 1, otherwise 0. Finally, F and V are concatenated and classified by an LVQ neural network, which output is integrated by a temporal layer named time potentials in order to reduce misclassifications. In a test course with a scooter approaching the reference car at 30km/h, the system could detect the approach 4.7s before the arrival. In normal city traffic, approaching vehicles could be detected 7.1s before the arrival. The accuracy is higher than 95% for less than 3s and around 88% for 4s. Only vehicles separated by more than 5 seconds from other vehicles and in the same or immediate neighbor lanes were considered. The system was implemented in an Altera Stratix II device. The aforementioned blocks used, respectively, 8242, 358 and 394 ALUTs and around 170kbits of memory. Multiples vehicles still presents a challenge and are the main focus of the future steps of this research.