|
Aiming at the actual traffic driving environment, this paper proposes an acoustic perception method of intelligent vehicle based on microphone array, and combines the visual perception method of monocular camera to study the acoustic visual fusion perception method. The main research contents are as follows: Firstly, an acoustic simulation environment for intelligent vehicles is designed. Secondly, a sound source localization algorithm based on time delay estimation is proposed. Based on the geometry of the microphone array, the sound source localization problem is abstracted as a time delay estimation problem, and the sound source localization is realized. The received signal of the microphone is added by weighted delay to form a beam with a certain direction, which is visualized as a spatial SRP power map. The passive sound reception is transformed into an active search for space, which improves the positioning accuracy in low SNR and reverberation environment. Thirdly, a sound source perception localization algorithm combining deep learning network and delay algorithm is developed. Based on SELDnet neural network, the SRP power map is used as input, and the cyclic convolution layer is replaced by the temporal convolution layer. The simulation results show that the positioning accuracy of the fusion positioning network is 65.04 % higher than that of the SELDnet network in the ideal environment and 76.19 % higher in the low SNR environment. The feasibility verification experiment and simulation experiment of the algorithm are carried out, and the sensor data in the real vehicle environment is collected for the real vehicle data verification experiment. The results show that the average positioning error of the acoustic vision fusion algorithm is 0.1 m at the 60 m scale. At the same time, the algorithm also has strong adaptability to visual blind spots, which proves the deployment potential of the algorithm in the actual environment.
|
