1. Introduction

Long-range light detection and ranging (lidar) of active illumination optical imaging possesses many advantages, such as high-resolution, anti-environmental interference, and all-time operation. These merits allow long-range active optical imaging to have widespread applications, ranging from remote sensing [1], satellite-based global topography [2], and airborne surveillance [3] to target recognition and identification [4]. Nowadays, active illumination optical imaging lidar can be generally classified into two broad categories with the imaging device: array imaging and point-detector-based imaging by computational reconstruction. The array imaging lidar is relatively mature and widely applied in the visible light band [5,6]. It generally employs a pulse laser to capture objects at a specific distance via time-of-flight (TOF). In non-visible bands, e.g., ultraviolet, infrared, or terahertz, array imaging is limited thanks to the array detectors with high resolution are usually unavailable. On the other hand, point-detector-based imaging may have more application potential due to its broader spectrum range and higher quantum efficiency, especially in non-visible light as the traditional array imaging is incapable. As one of the representative imaging technologies with the single-pixel detector, point by point imaging lidar is restricted by its small divergence angle and receiving field of view, the resolution, efficiency, sample preservation, and signal-to-noise ratio are difficult to optimize simultaneously [7,8].

Single-pixel imaging (SPI) [9,10] is another representative imaging strategy with the single-pixel detector, which is often referred to ghost imaging (GI) [11–15]. SPI recovers the image information by analyzing the correlations between the predetermined illumination patterns by the spatial light modulator (SLM) and measured intensities by the single-pixel detector. SPI has been demonstrated in many non-visible bands imaging, such as infrared [16], terahertz [17], X-ray [18]. Generally, to obtain an N-pixel image for the SPI, one needs at least M = N measurements to meet β=M/N = 100%, where β represents the sampling ratio. As can be seen, the higher the imaging resolution, the more measurements are required. How to realize high-quality imaging in the under-sampling that needs to be overcome for practical applications with SPI technique. Fortunately, by utilizing the sparsity of natural scenes, compressed sensing (CS) [19–21] and the orthogonal sub-sampling concept, such as the deterministic Hadamard [22], Fourier [10], and wavelet [23] base can be applied to recover high-quality images under sub-sampling. For example, via sparsity constraint, Gong W et al. [24] demonstrated a GI setup in 532 nm with a dual optical path structure. Experimental results show that 3D scenes can be reconstructed with global measurements below the Nyquist limit in the real atmospheric environment. To the best of our knowledge, in the dual optical path structure GI setup, there will inevitably be some uncertainty between the recorded light intensity distribution by the array detector and the truth. Besides, the array detectors often face an unavailable dilemma in some non-visible bands. Taking these into consideration, therefore, SPI lidar may have some advantages to imaging in the long-range real atmospheric environment.

For the long-range active illumination imaging lidar, one often expects a large FOV and high resolution simultaneously. However, for a certain imaging system, the FOV usually shall match the divergence angle of the transmitting laser. A large divergence angle inevitably lowers the energy density of the emitted laser, thus shortening the imaging distance. On the other hand, high-resolution imaging requires a small pixel size, which will lead to a low detection SNR. So, many pixels are usually binning together to improve SNR but at the cost of spatial resolution. Based on the above cognition, without sacrificing the corresponding imaging resolution of pixels, this paper demonstrates a scanning single-pixel imaging lidar (SSPIL) based on a commercial digital micromirror device (DMD). The SSPIL makes some trade-offs among imaging efficiency, resolution, FOV, divergence angle, and detected distance. Some experiments are carried out in the real urban atmospheric environment to assess the performance of the imaging setup. The results show that SSPIL has the ability of long range with high efficiency and resolution. The organization of the paper is as follows. The experimental setup is introduced in Section 2. And the imaging strategy of SSPIL is detailed in Section 3. Three experiments and the corresponding imaging results are presented in Section 4. Finally, Section 5 concludes the paper.

2. Experimental setup

Figure 1 shows the schematic diagram of the SSPIL experimental setup. A diode-pumped Nd: YAG laser of the center-wavelength λ=1064 nm is employed as the illuminator, its pulse width is less than 7 ns at a pulse repetition frequency of 400 Hz, and the maximum pulse energy is 400 mJ. A two-dimensional (2D) galvanometer is used to change the pointing direction of the output laser beams. The scanning range of the galvanometer is ±0.393 rad with a resolution of 12 µrad. A collimator lens is used to set the laser divergence angle to ∼0.1°×0.1°. After that, the squared laser beam irradiates the scene and interacts with the imaged targets. A self-made telescope with the aperture Φ=240 mm and the focal length f = 730 mm collects the photons back-reflected from the interaction, and then projected on a DMD (ViALUX DLP V-9601). The DMD has an active 1920 × 1200 mirror array area of 20.7 mm×13.5 mm, each of which tilts +12 degrees or −12 degrees with respect to the DMD surface forming binary patterns. Thus, the whole receiving FOV corresponding to the squared area of the employed DMD is about 18 mrad×18 mrad (or 1.03°×1.03°). An avalanche photodiode (APD) is attached to the +12 degrees sides of the DMD. To match the laser divergence angle but without sacrificing the spatial resolution corresponding to one micromirror pitch of the DMD, here we divide the 1200 × 1200 pixels of the central squared region of the DMD into 100 equal parts, allowing each part containing valid pixels of 128 × 128 corresponding to a sub-FOV with ∼0.1°×0.1° (Note that the 128 × 128 pixels of the last row and the last column overlap with the corresponding adjacent row and column). That is, only 128 × 128 pixels of the sub-FOV are used for imaging, and the other pixels is set to 0. Thanks to the APD being fixed on the +12 degrees sides of the DMD, this operation can suppress most invalid stray light due to most background light being reflected to the −12 degrees sides of the DMD.

 

Fig. 1. Schematic diagram of the SSPIL experimental setup. DMD: Digital Micromirror Device. A and B indicate the different laser beams pointing controlled by the galvanometer.

Download Full Size | PDF

A high-speed digitizer of 400MHz is used in our setup, which can reach a spatial resolution of 0.357 m. A program written with C/C++ in Microsoft Visual Studio 2013 development environment is designed to control the DMD, galvanometer, acquisition, etc. The DMD is set in the slave mode, which is triggered by the Q-switching synchronization signal from the laser. And the output synchronization signal from the DMD triggers the digitizer to realize synchronous data acquisition. Limited by the repetition rate of the laser source, the refresh frequency of the DMD is set to 400 Hz in our subsequent experiments.

3. Scanning single-pixel imaging strategy

Concerning modulation patterns, the deterministic orthogonal Hadamard base which is derived from the Hadamard matrix is employed in our experiments. Further, considerations of the limitation of the repetition frequency and making a trade-off the image reconstruction quality and acquisition time, the first 25% Hadamard basis patterns in the ‘Russian Dolls’ order [25] are applied to our imaging experiments. Thus, the number of patterns is 8192 (considering differential measurement) for the imaging resolution of 128 × 128. The time consumption for imaging a 128 × 128-pixle image in one modulation and measurements circle is about 20.48 s at a 400 Hz refresh rate.

Unlike the conventional point-by-point scanning imaging and SPI lidar, the strategy of SSPIL is divided into two steps: scanning search and staring imaging. The aim of scanning search is to sense which sub-regions or sub-FOVs of DMD exist targets possibly, and the staring imaging is to image the specific sub-FOVs where targets exist. The processes are demonstrated schematically in Fig. 2, the upper (Fig. 2(a)-(e)) is corresponding to the scanning search, and the lower (Fig. 2(f)-(g)) is for the staring imaging.

 

Fig. 2. Strategies of SSPIL. The process of SSPIL is divided into two steps: scanning search and staring imaging. (a) are the scanning search patterns which are composed of 100 patterns, the valid pixels of per pattern are 128 × 128, the pixels at those positions are 1, and the rest are 0. (b) is demonstrated the scanning search for the background areas. (c) are the time-resolved measurements corresponding to background region. Due to no object illuminated by the squared laser beam, the measurements are relatively flat. (d) is demonstrated the scanning search for the imaging objects which are illuminated by the squared laser beam. Since the scanning search pattern is switched synchronously, the corresponding time-resolved measurements (e) at this time emerge jumps at some ranges. (f) are the staring imaging patterns which are composed of 8192 patterns based on Hadamard base. (g) is the reconstructed image by combining the measurements with the corresponding imaging pattern.

Download Full Size | PDF

As noted, before starting the experiments, a critical operation need conduct to calibrate the scanning parameters of the 2D galvanometer and the imaging sub-FOVs of the DMD. The aim is to ensure the position of the laser beam pointing by the 2D galvanometer and the certain sub-FOVs interrelated exactly for scanning search and staring imaging. As mentioned above, to match the 0.1°×0.1° divergence angle of the outgoing laser, in scanning search process, we produce 100 patterns with 1920 × 1200-pixel, but the valid pixels per pattern is 128 × 128, as shown in Fig. 2(a). The pixels at those positions are 1, and the rest are 0. We name them scanning patterns. The galvanometer sequentially scans the spatial orientation of the laser beam, e.g., while the active light spot illuminates the scene demonstrated in Fig. 2(b), the corresponding scanning pattern on the DMD is the first one shown in Fig. 2(a), the corresponding time-resolved measurements are relatively flat due to no object existing, as shown in Fig. 2(c). Consequently, we consider this region (128 × 128-pixel) of the DMD as the background scene. Next, while the active light spot illuminates the area demonstrated in Fig. 2(d), the corresponding scanning search pattern (e.g., the 59th in Fig. 2(a)) is simultaneously switched. At this time, the corresponding time-resolved detected echo intensities would have some variations along with distances, as demonstrated in Fig. 2(e). There must be existing “something” in this position illuminated by the laser spot. Then, the scanning search will be suspended and switched to staring imaging. In staring imaging, the light pointing holds invariably. The operation mode and reconstruction method of the staring imaging are the same to the traditional TOF SPI lidar. Here, the staring imaging patterns (e.g., Fig. 2(f)) P_i (x, y) (i = 1, 2, …, 8192, x and y represent the spatial position of the patterns, both x and y are equal to 128) of this region will be loaded in the DMD. The corresponding time-resolved detected signal of the imaged scene can be expressed as follows:

Here, U_i(d) are the time-resolved measurements corresponding to the certain sample time t, so the distance d = ct/2 (c is the light speed). The f(x,y;d) represent the object images at different distance d. For imaging reconstruction, the linear iterative algorithm is applied to recover the image by combining the detected intensities with the corresponding staring imaging patterns:

In addition, imaging in the real atmospheric environment, the time-resolved detected intensities unavoidably mix with the background noise, electronic noise, etc. Besides, atmospheric turbulence also may affect the imaging quality. So, as a comparison, an optimization algorithm of TVAL3 [26] is also applied to recover the image under sub-sampling in our experiments.

4. Experiments and analysis

Three imaging experiments with SSPIL were conducted. The laser pulse energy is set at 10 mJ. The first imaged target is a tower crane for the building about 3.9 km away from the SSPIL, as shown in Fig. 3(a). The sampling points are set to 32000, the maximum detected distance is up to 12 km under the 400 MHz sampling ratio. The laser beam pointing is changed sequentially by the 2D galvanometer, the 100 scanning patterns corresponding sub-FOVs of the DMD are changed simultaneously. Through scanning search, the time-resolved measurements corresponding to 10 sub-FOVs (the 43rd, 52nd, 53rd, 54th, 55th, 56th, 57th, 58th, 59th, and the 63rd demonstrated in Fig. 3(b)) emerge jumps around 3.9 km (Fig. 3(c)), respectively. As mentioned above, the staring imaging will be carried out in these ten regions. We know the time-resolved measurements corresponding to the first imaging pattern (Hadamard base) are bigger than the rest, so we first obtain the index of the maximum echo signals corresponding to the first pattern. Thus, the measurements corresponding to all other imaging patterns can be obtained with the index. Ten images can be recovered using the linear iterative algorithm. And then, arranging these ten images according to the employed sub-FOVs regions of the DMD, a picture of 1024 × 384-pixel can be obtained, as shown in Fig. 3(d). As can be seen, the hoisting ropes, the shape, and the structure of the tower crane are unfolded. In SSPIL, the time consumption for the scanning search is almost negligible; the time consumption for the 1024 × 384-pixel imaging is 10 × 8192/400 = 204.8 s. As a comparison, the time for only scanning and detection of the traditional point-by-point scanning imaging lidar is about 1024 × 384/400 = 983.04 s. Hence, the proposed SSPIL reduces almost 80% of the time consumption compared with traditional point-by-point scanning imaging. As can be seen, the SNR of the image obtained has a little poor within only one modulation and detection cycle. Given the efficient denoising ability of compressed sensing, we use TVAL3 to reconstruct the images by combining the measurements and the Hadamard base employed, and the results are shown in Fig. 3(e). The images are better than that with the iterative algorithm. Generally, multiple acquisitions can depress the random noise of measurements. We demonstrate the imaging results are recovered by applying the iterative algorithm and TVAL3 in 20 modulation and detection cycles, shown in Fig. 3(f) and Fig. 3(g), respectively. Compared with the results shown in Fig. 3(d), imaging quality is significantly improved. But the imaging quality recovered by TVAL3 is slightly improved, unlike the images retrieved have a lot of enhancement in one modulation and detection cycle.

 

Fig. 3. Imaging with SSPIL for a ∼3.9 km tower crane. (a) is the imaging scene, and the inserts are the imaged target and the experimental setup captured by a camera. (b) is demonstrated the 10 imaging sub-FOVs of the DMD. (c) are the time-resolved measurements corresponding to one of the 10 imaging sub-FOVs of the DMD, the insert in the upper right corner shows a closer view of the time-resolved measurements corresponding to the imaged objects. (d) and (e) are the reconstructions with the iterative algorithm and TVAL3 in one projection and detection cycle, respectively. (f) and (g) are the reconstructions with the iterative algorithm and TVAL3 in 20 modulation and detection cycles, respectively.

Download Full Size | PDF

The second imaging target is a telecommunication transmission tower about 7 km from the imaging setup, as shown in Fig. 4(a). The scanning search and imaging processes are the same as for the first imaging experiments. Through scanning search, the time-resolved measurements corresponding to 7 sub-FOVs (the 52nd, 53rd, 54th, 55th, 56th, 57th, and 58th) emerge jumps around 7 km, respectively. Seven images can be obtained by staring imaging. And then, arranging these seven images according to the employed sub-FOVs regions of the DMD, an 896 × 128-pixel image can be obtained. Figure. 4(b) and Fig. 4(c) demonstrate the results in one modulation and detection cycle by applying the iterative algorithm and TVAL3, respectively. The imaging quality of the farther telecommunication transmission tower has a certain degradation. Figure. 4(d) and Fig. 4(e) demonstrate the results in 30 detection cycles via the iterative algorithm and TVAL3, respectively. As can be seen, the imaging quality has a considerable improvement compared to the results of Fig. 4(b) and Fig. 4(c).

 

Fig. 4. Imaging with SSPIL for a ∼7 km target. (a) is the imaged scene, and the insert in the lower left corner shows the imaged target object captured by a camera. (b) and (c) are the reconstructions with the iterative algorithm and TVAL3 in one projection and detection cycle, respectively. (d) and (e) are the reconstructions with the iterative algorithm and TVAL3 in 30 modulation and detection cycles, respectively.

Download Full Size | PDF

TOF active illumination SPI both can obtain the reflectivity and depth map of the imaging scene [27] and is a very promising tool for object recognition and remote sensing. Next, we conduct a time-resolved imaging experiment for different distance targets with SSPIL. A small tower adjacent to the tall telecommunication transmission tower (as shown in Fig. 4(a)) at a distance of about 7 km away the imaging setup is very suitable for us to complete such an interesting experiment. Some plants are in front of the tower, causing the latter to be partially covered, as shown in Fig. 5(a). The plants and the tower fall in a sub-FOV of DMD. The corresponding time-resolved detected echo signal are shown in Fig. 5(b). The insert of Fig. 5(b) shows 120 measurements corresponding to the imaging targets. Figure 5(d) illustrates the image cube recovered by applying the iterative algorithm for the 120 adjacent measurements. Note that there are forty images in the intermediate transition position omitted in Fig. 5(d). As can be seen, the time-resolved imaging has well revealed the sequence of interaction between imaging targets and illumination pulse laser beam. The scene reflectivity (see Fig. 5(c)) is calculated by averaging the image slices. In the image cube, each transverse pixel (x, y) has an intensity distribution, so the depth map of the imaged scene (see Fig. 5(e)) can be subsequently obtained by getting the maximum in these transverse pixels along with the distance/time. As seen, the depth map further reveals the two-dimensional and longitudinal range information about the imaging targets. It should be noted that the recovered results are obtained in 25 modulation and detection cycles.

 

Fig. 5. Three-dimensional imaging with SSPIL. (a) Demonstration of the imaging scene, and the insert is the target captured by a camera. (b) The time-resolved measurements correspond to different distance objects, the insert in the upper right corner shows a closer view of the 120 time-resolved measurements corresponding to 45 m range. (c) and (e) are illustrated the reflectivity and depth map of the imaged scene, respectively. (d) is the image cube at a different distance with the 120 adjacent measurements, 40 images in the intermediate transition position are omitted and represented with the three red dots.

Download Full Size | PDF

5. Conclusion and discussion

In summary, we demonstrate an active illumination SSPIL. A 2D galvanometer is employed in our setup to scan the scene illuminated by the output laser spot sequentially. To make trade-offs among the imaging efficiency, imaging resolution, and detected distance, we divide the whole imaging regions of the DMD into many sub-regions or sub-FOVs to match the low divergence angle of the output laser beam. Unlike the conventional point-by-point scanning imaging and SPI lidar, here the imaging strategy of SSPIL is divided into scanning search and staring imaging processes. These strategies save time consumption for imaging background areas and thus to improve imaging efficiency. The imaging experiments for three targets at different distances were conducted in real urban atmospheric conditions to evaluate the performance of the setup. The preliminary results show that SSPIL has the ability for long-range imaging with high efficiency (compared with pointing-by-pointing scanning imaging) and resolution. Also, from the imaging results, we found that multiple samples can improve the SNR of imaging in the real atmosphere.

The superiority of SSPIL is in sensing the targets in the long-range sparse scenes. Limited by the position of the setup located, further imaging objects for testing the performance of SSPIL will be conducted in our next work. As the introduction mentioned, the long-range active illumination imaging lidar usually expects to have a large FOV and high resolution, we think the present work may provide a valuable alternative way to achieve these requirements.