Objective The laser distance measurement system is widely used in various fields such as long-distance measurement, atmospheric concentration monitoring, three-dimensional laser imaging, and star-to-ground laser ranging, due to its small volume, high accuracy, and fast response speed. The echo time of the time-of-flight method is the focus of this technology. The situation of “strong background noise and weak signal light intensity” is that the laser ranging technology needs to face. On one hand, the background noise is too strong relative to the signal light and the signal-to-noise ratio is low, which makes signal processing difficult. On the other hand, the avalanche photon diode (APD) unit with Geiger mode is prone to channel saturation or channel blockage under strong background noise. This paper derives and discusses in detail the spatial three-dimensional convolutional neural network signal processing model, which combines the photon counting method with the convolutional neural network method. The simulation verifies that this algorithm can effectively identify the weak signal light from the strong background light.
Methods A theoretical model is established via theoretical derivation and subsequently it is verified via simulation. Firstly, different from the previous laser ranging signal processing that only considers the probability distribution of the signal response in time domain and presents the Poisson distribution, this article adds a discussion of the probability distribution in spatial domain on this basis. We believes that the signal light and background light received by the receiving lens barrel used for laser ranging have spatial distribution differences. Therefore, in the simulation, the signal light is set as the far-field spot that obeys the two-dimensional Gaussian-like distribution, and the background noise is set to be evenly distributed in spatial domain. Then, this article simulates the insertion of short signal light into long background noise, and obtains the original signal used in this algorithm with the Geiger mode 4×4 APD detection. Second, this algorithm is designed to process the spatial three-dimensional convolutional neural network signal that combines the photon number counting method in the one-dimensional time domain with the convolutional neural network in the two-dimensional spatial domain. Finally, by comparing multiple sets of simulated signals and drawing conclusions through the recognition of simulated signals, it is verified that the proposed algorithm has an improved effect on the recognition of signal light.
Results and Discussions The simulation results show that it is theoretically feasible to process the Geiger pattern array APD signal through the spatial convolution method, and the signal recognition can be nearly doubled. In the article, Figs. 10 and 11 respectively show the fitted signal obtained by simple operation and that after using 2D convolution treatment. The comparison shows that this algorithm can hide the noise in the noise. The array fitting signal is effectively strengthened to highlight the spikes and effectively improve the signal recognition. Figure 12 shows the signal diagrams fitted by using four 3×3 feature convolution kernels. Obviously, after the same 4×4 array signal is processed with different N-dimensional convolution kernels, the signal recognition is somewhat different. But the peak point is close. Figure 13 and 14 respectively show the fitted signal diagrams and signal recognition degree curves after using different N×N feature convolution kernels. The results show that as the dimensionality of the convolution kernel increases, the signal recognition first increases significantly. The big follow-up tends to be flat. Obviously, when the dimensionality of the detection signal and that of the signal processing convolution kernel matches, the signal recognition can reach the best. After using the Geiger pattern array detection device, the introduction of spatial domain signal processing method is beneficial to signal processing.
Conclusions In this paper, a three-level signal recognition optimization algorithm based on photon counting and convolutional neural network is designed by combining the existing single-photon detector array with its auxiliary circuit used in laser ranging and based on the idea of multi-level improvement of signal recognition. This algorithm not only inherits some of the calculation methods of the traditional time-of-flight algorithm, but also combines the spatial statistical distribution characteristics of the photons scattered on the plane of the two-dimensional detection array, and comprehensively considers the traditional one-dimensional signal processing method and the two-dimensional convolutional neural network. The new idea constitutes a spatial three-dimensional convolutional neural network program for processing laser ranging signals. After simulation verification, the program is feasible in theory as well as engineering.
