To address the issue of high miss rates for distant small objects and the diminished system detection performance due to the influence of hazy when autonomous vehicles operate on mountain highways. We propose a framework for small object vehicle detection in hazy traffic environments (SHTDet). This framework aims to enhance small object detection for autonomous driving under hazy conditions on mountainous motorways. Specifically, to restore the clarity of hazy images, we designed an image enhancement (IE), and its parameters are predicted by a convolutional neural network [filter parameter estimation (FPE)]. In addition, to enhance the detection accuracy of small objects, we introduce a cascaded sparse query (CSQ) mechanism, which effectively utilizes high-resolution features while maintaining fast detection speed. We jointly optimize the IE and the detection network (CSQ-FCOS) in an end-to-end manner, ensuring that FPE module can learn a suitable IE. Our proposed SHTDet method is adept at adaptively handling sunny and hazy conditions. Extensive experiments demonstrate the efficacy of the SHTDet method in detecting small objects on hazy sections of mountain highways.

Transport of intensity (TI) is a well-known non-interferometric technique for phase retrieval. The TI and phase equations result from the Helmholtz equation and show the coupling of intensity and phase during optical propagation. TI is an alternative to digital holography, which requires a reference beam for a recording of the interference pattern. However, the conventional TI method has an experimental challenge in that mechanical displacement of the camera or object is needed to record the optical intensities at multiple defocused planes, which can cause errors from misalignments. This work expands on a modified TI technique that avoids mechanical displacements, instead invoking the use of electrooptic materials to create an optical phase difference and hence optical path length through the application of a bias voltage. We demonstrate the use of the modified TI equation (TIE) through simulation and experiment by selecting suitable objects and a biased nematic liquid crystal cell made from pentyl-4-cyanobiphenyl (5CB). The corresponding modified transport of phase equation is also derived and is used to enhance the accuracy of the modified TIE. After providing simulation results for imaged phase retrieval, we demonstrate the unwrapped image phase and hence height or profile extraction for an object with three-dimensional topography using this technique.

Many stitching studies employ improved scale-invariant feature transform (SIFT) or speeded up robust features (SURFs) stitching algorithms, owing to their excellent robustness against variations in scale and rotations of input images. Traditional SIFT-based stitching algorithms suffer from the limitation of having insufficient feature points that can be extracted and matched from untextured and smooth objects, resulting in stitching failure. To overcome this issue, we propose a system that uses a proposed laser pattern projection (LASPP)-pasted texture technique before feature point extraction and description. The proposed system can extract and describe feature points based on SIFT feature extraction techniques. The progressive sample consistency (PROSAC) algorithm was used to remove false matching points and enhance the accuracy in matching feature points. The experimental results clearly revealed the effect of the pattern projected using LASPP on the stitching results. Furthermore, the efficiency of the proposed system, which employed the PROSAC algorithm, in matching feature points, was compared with that of the traditional system, which used the random sample consensus algorithm. Additionally, the stitching quality of the proposed system was compared with that of other SIFT- and SURF-based stitching systems.

Spatial aliasing in its most pronounced form is seen as a Moiré pattern in (sampled) images. Less dramatically, aliasing is a form of image quality (IQ) degradation and exists at some level within any (sampled) image. More sampling tends to improve IQ due to less aliasing; however, there are drawbacks. Spatial aliasing has been recently quantified by Mudge [Appl. Opt. 62(13), 3260–3264 (2023)] for imaging sensors (optics plus detector). This quantification allows a trade to be made between the acceptable aliasing errors imbedded within the (sampled) image and the penalty, or cost, associated with additional sampling, e.g., increased complexity, data throughput and storage, and reduced signal-to-noise ratio for fixed arrays or increased scan time. In this work, we examine several existing and useful imaging sensors along with their imagery and aliasing errors to appreciate how well these existing systems are designed with respect to sampling to better inform how future systems could potentially be improved. Finally, from these analyses, a 2% aliasing error rule is extracted initiating a universal aliasing boundary.

The cutting-edge research in high-dynamic-range (HDR) display, such as inverse tone-mapping operators for HDR TV, HDR image quality (IQ) evaluation, and HDR image enhancement, suffer from a lack of reliable training and testing images. To address this problem, we combined the HDR imaging technique with high-accuracy spectral images to construct a first-of-its-kind HDR spectral radiance image. A framework was presented to capture multi-exposure hyperspectral images of HDR outdoor scenes and establish the HDR spectral image. By employing spectrophotometric calibration, the spectral radiance-based HDR images were created. The metadata of the spectral images provide crucial colorimetry information that was absent in the existing datasets. To validate the superiority of the proposed HDR spectral images in typical applications, the spectral sensitivity functions of a camera equipped with a multispectral-filter-array sensor was optimized, and the IQ evaluation results indicate a better adaptability over existing ones. Furthermore, the HDR image augmentation suggests that it is possible to provide a large number of RGB training images for various research objectives based on the proposed images.

Focus-tunable liquid lenses are used in various applications due to their compact size, light weight, low power consumption, and cost effectiveness. They have the potential for use in space applications, such as focus compensation, optical communications, and imaging systems. However, liquid lenses have not yet been evaluated for use in the space environment. This work focuses on characterizing operational differences of commercially available liquid lenses from Corning Varioptic and Optotune between Earth gravity, microgravity, and hypergravity environments. Results show a linear drift in the tip/tilt of 0.80 and 4.20 mrad going from 1 to 0 g for the Corning Varioptic A-39N0 lens and Optotune EL-16-40-TC-VIS lens, respectively, with lower optical aberrations in microgravity. Additionally, a significant but small increase in focal power going from 1 to 0 g by 0.02 D is observed for the Optotune lens. No significant change in focal power is observed for the Corning Varioptic lens tested in this experiment. Additionally, potential multi-beam interference is observed in defocus patterns of the Corning Varioptic lens tested during the experiment.

One of the primary activities in emissive infrared imager design is the trade on whether to use midwave infrared (MWIR) or longwave infrared (LWIR) in the application. Applications include target acquisition (both target search and target identification), threat warning, aircraft detection, and pilotage. There has been a great deal of work in the characterization of MWIR versus LWIR target signatures. There has been much less work in the characterization of scene (sometimes called background) contrast. The scene contrast of the background can be just as important in the performance of the sensor in an application. A few examples are: (1) a high scene contrast with high clutter contrast can make target search much more difficult, (2) a high scene contrast with image-based navigation can enhance the performance of location estimation, and (3) high scene contrast with mobility sensors can enhance the flying performance of a rotorcraft pilotage system. We discuss the differences observed in scene contrast between MWIR and LWIR bands. This provides a scene contrast characterization for emissive infrared applications. Radiometrically calibrated imagery is acquired with MWIR and LWIR cameras in various environments, and the measured MWIR and LWIR scene contrast is compared. The radiometric comparison is performed in terms of the standard deviation of the scene equivalent blackbody temperature. Comparisons are provided under different conditions, such as rural versus urban and day versus night. This comparison enables the infrared system designer with the means to perform detailed engineering trades.

Underwater images often suffer from problems such as low contrast, color distortion, and blurred details, which have a negative impact on subsequent image processing tasks. To mitigate such problems, we propose an algorithm that combines an underwater physical imaging model with a convolutional neural network. The physical imaging model has two important types of parameters: background scattering parameters and direct transmission parameters. For the background scattering parameters, we divide them into three levels, which are primary information, secondary information, and advanced information, and design three subnetworks for feature extraction to represent them. For the direct transmission parameters, we decompose them into two levels, which are shallow transmission information and deep transmission information, and also design two subnetworks to indicate them. Experimental results show that, compared with other enhancement algorithms, the proposed algorithm can not only effectively correct the color deviation and enhance the object edges and texture details but also can obtain superior values of objective evaluation metrics in terms of peak signal-to-noise ratio, structural similarity, underwater color image quality evaluation, and underwater image quality measure.

Transverse spatial structure associated to complex vector vortex beams (VVBs) is classified through an image texture approach, which is based on Haralick features derived from an azimuthally sensitive gray-level co-occurrence matrix (a-GLCM). Previously, the proposed texture identifier has been tested on azimuthally symmetric digital images presenting an improvement in comparison with conventionally constructed GLCM. Then, considering their suitable spatial symmetry, Laguerre–Gauss mode vector beams were numerically generated to characterize its transverse section. Nonetheless, because of the complex structural information, VVBs are decoupled in three 8-bit grayscale images (channels) corresponding to intensity, orientation, and ellipticity parameters. a-GLCMs and Haralick features for each channel are computed with the aim of quantifying transverse structure evolution as a function of topological charge. Also, image scaling effects are studied to test the approach robustness. Even though it has simplicity, texture analysis based on a-GLCM provides an insight to classify VVB structure.

We comprehensively investigated the influence of laser wavelength on the detection distance at different clear air turbulence (CAT) intensities, distance resolutions, and Brunt–Väisälä frequency theoretically. It is found that there exists the optimal wavelength for maximal detection distance. When the distance resolution is fixed, the detection distance of 532 nm LiDAR is further compared with the 355 and 1064 nm counterparts and is more suitable under moderate to severe CAT. Our findings break the traditional cognition that “ultraviolet wavelength LiDAR is the most suitable for CAT detection” and may promote the remote detection research of CAT.

Robotic grippers combined with visual-touch sensor fusion have been widely used in the manufacturing industry. However, due to limited control of the force applied to the mechanical gripper, it is hard to achieve stable object grasping. Therefore, to control the force of gripping objects, we propose a robotic rotary gripper fused with light detection and ranging (LiDAR)-tactile. We use LiDAR to obtain distance information and fiber Bragg grating as the tactile sensor for force feedback. In order to ensure the low-loss transmission of the optical signal in the rotation process, the fiber optic rotary joint is employed and assembled into the system. Finally, the experimental results show that our proposed system can achieve a stable grasp of objects with different shapes by precisely controlling the force, which can meet the demands for precise robotic grasping and fast manipulation.

In fringe projection profilometry, applying pre-distortion to fringe patterns reduces the errors caused by projector lens distortion. However, it is important to note that discontinuous fringe patterns, such as binary fringe patterns, introduce additional errors when using pre-distortion methods. While post-undistortion methods are applicable for discontinuous fringe patterns, the computation is typically time-consuming. We propose a linear-grid model for correcting lens distortion. First, we select multiple equidistant points within the grid to calculate the linear parameters and store them as look-up tables (LUTs). Second, by rounding down the captured distorted point to the nearest integer point, we obtain the index value for LUTs. Finally, we achieve real-time compensation for distortion error through linear expressions. The experimental results show that the proposed effectively mitigates the distortion by a factor of 6 × in terms of root mean squared error. Additionally, it exhibits a computational speed of 409.50 fps, which is an improvement compared to the traditional iterative model at 39.48 fps and the scale-offset model at 264.48 fps.

Multispectral light detection and ranging (LiDAR) is an emerging active remote sensing technique that combines distance and spectroscopy measurements. The reflectance spectrum is known to enable material classification. However, the spectrum also depends on other surface parameters, particularly roughness. Herein, we propose an extension of multispectral to polarimetric multispectral LiDAR and introduce unpolarized and linearly polarized reflectance spectra as additional features for classifying materials and roughness. Using a bench-top prototype instrument, we demonstrate the feasibility and benefit of acquiring unpolarized and linearly polarized reflectance spectra. We analyze and interpret the spectra obtained with two different spectral resolutions (10 and 40 nm) from measurements on test specimens consisting of five different materials with two different levels of surface roughness. Using a linear support vector machine, we demonstrate the potential of the different features for enabling material and roughness classification. We find that the unpolarized reflectance spectrum is well suited for classifying materials, and the linearly polarized one for classifying roughness. In both cases, the performance is much better than using a standard reflectance spectrum offered by multispectral LiDAR. We identify polarimetric multispectral LiDAR as a technology that may significantly enhance surface and material probing capabilities for remote sensing applications.

The self-rotating three-dimensional (3D) visual scanning system is a prevalent technique for 3D reconstruction. However, the misalignment of the scanner’s rotation axis with the camera’s origin often necessitates specialized calibration tools to establish their relative positions. We propose a rotating axis calibration method for the rotation axis, which eliminates the need for specific tools. A rotating mechanism is used to rotate the 3D camera to different angles and carry out 3D reconstruction of the same object with curved surface characteristics, respectively. The resulting sequence of 3D point clouds is registered to derive the camera’s position transformation matrix corresponding to each rotation angle. Subsequently, we compute the positional relationship between the camera and the rotation axis by solving the equation. To prevent excessively large rotation angles that would cause the object beyond the camera’s field of view, the camera field extension method is proposed, which enhances equation redundancy and leads to improved calibration accuracy. Experiments prove the flexibility and accuracy of the proposed method.

A silica-based 90 deg optical hybrid based on a 4 × 4 multimode interference coupler with a refractive index contrast of 0.36% is designed and fabricated for coherent receiver. The measured results of this device show that the common mode rejection ratios are better than 23 dB and excess loss including a coupling loss of −0.3 dB is less than −0.9 dB in C band (1530 to 1565 nm). The phase error is less than ±4 deg. The designed low-cost 90 deg optical hybrid is promising for further improving the sensitivity of quadrature phase shift keying coherent receivers.

This work presents the exact analytical formula for calculation of the longitudinal chromatic aberration of a lens, in contrast with the approximate nature of the traditional approach based only on the dispersion of the optical material. This formula additionally takes into account the radii of curvature of the lens surfaces, lens thickness, object and image distances, and the chromatic aberration of the previous optical system. The presented formula is used for the derivation of rigorous equations for color correction of simple lens systems. To test them, this work presents examples of achromatic and apochromatic lenses that are designed by solving these equations. These results prove the usefulness of the presented approach for color correction and open the way for a broad and rigorous study of chromatic aberrations in various optical systems.

Lithium niobate thin film (TFLN) electro-optic modulators (EOM) are widely used in optical communication systems due to their high bandwidth, low half-wave voltage, and low energy loss. Based on the analysis of the finite element method, we optimize the structure parameters of the modulator and propose an EOM with an S-bend coupler and an electrode gap covered with a SiO₂ layer. The device exhibited a half-wave voltage-length product of 2.13 V · F, a bandwidth greater than 75 GHz, and an optical loss <0.1 dB / cm. Besides, we also propose employing the linear regression algorithm in machine learning to help determine the parameters of high-performance modulators, which provides a new idea for the design and optimization of TFLN modulators.

Improvement of the quality factor has now been an active area of research due to the growing interest of researchers in the field of the photonic crystal-based ring resonator and its applications in photonic devices. The introduction of randomness into the photonic crystal structures has proved to be an effective way to enhance the quality factor. A theoretical model of photonic crystal structure with a size disorder of the rod radius is investigated. The radius of the inner rods of the ring is modulated with the Gaussian random distribution function. The effect of a change in the disorder parameter (standard deviation) on the quality factor of the photonic crystal structure and the effects on the quality factors for variation of the refractive index of the structure and the analyte sample are evaluated.

The preparation of a non-absorption window (NAW) through impurity-free vacancy disordering (IFVD) induced quantum well intermixing (QWI) in 915 nm GaInAsP/GaAsP semiconductor lasers is aimed at enhancing the catastrophic optical mirror damage (COMD) threshold and output power of semiconductor lasers. A process for the fabrication of high-power semiconductor lasers has been explored through the intermixing study of primary epitaxial wafers. In the experimental procedure, the epitaxial layers at the front and rear facets were first selectively etched, followed by the deposition of a 200 nm SiO₂ dielectric layer on their surfaces. Additionally, a 100 nm TiO₂ dielectric layer is applied to suppress QWI in the gain emission region. A NAW is achieved through a rapid thermal annealing process at 830°C for 180 s under atmospheric pressure. In the wafer-scale fabrication process, wide-stripe single-tube devices are produced with a light-emitting region width of 200 μm and a cavity length of 4000 μm. Test results reveal that conventional single-tube devices typically experience COMD at an input current of ∼20 A, with power output at around 19 W. In contrast, the single-tube devices with the NAW reach a continuous output power of 27.7 W at an input current of 30 A, representing a power enhancement of ∼45.8 % . No optical catastrophic damage is observed, and characteristics, such as slope efficiency and threshold current, remain unaltered. This research demonstrates that IFVD combined with the wafer-scale fabrication can effectively enhance the COMD threshold of 915 nm GaInAsP/GaAsP semiconductor laser devices, offering valuable insights into the study of high-power semiconductor lasers.

The accuracy of thermal analysis modeling for direct energy deposition is compromised if the Marangoni effect is ignored. The anisotropic enhanced thermal conductivity approach can improve the fidelity of the transient finite element (FE) thermal model, but the determination of the enhanced factors is generally based on empirical knowledge or trial and error. In this study, single track cladding experiments were conducted at different process parameters, the cladding process was recorded by an infrared camera, and the melt pool width and length were extracted from the infrared images. The best match enhanced factors were confirmed according to the comparison between the simulated and experimental melt pool sizes. The findings suggest that the best match enhanced factors are not a constant value, but rather vary with changes of process parameters. With the enhanced factors increasing, the peak melt pool temperature drops sharply at first and then tends to be stable. The melt pool width remains almost unchanged, whereas the melt pool length exhibits a slight increase. Laser specific energy was found to be a suitable metric to characterize the convection features and intensity of the melt pool. An empirical model of laser specific energy and enhanced factors was derived through linear regression analysis, which greatly simplified the determination of the enhanced factors. The experimental results showed that the error of the FE thermal model in predicting the melt pool size was less than 6.3%.

Exploiting the phase transition of vanadium dioxide (VO₂), a switchable terahertz multifunctional device integrating narrowband absorption and broadband polarization switching was proposed in our study. The transformation of VO₂ from insulating state to metallic state can realize different functions. When VO₂ was in metal state, the metamaterial device exhibited an absorption effect exceeding 90% in the range of 1.96 to 4.06 THz, and its amplitude was tunable via changing the conductivity of VO₂. Further, the modulation depth reached 171%. This device was also used as a broadband cross-polarization converter, achieving more than 97% cross-polarization conversion in the range of 2.03 to 3.93 THz. When VO₂ was in insulating state, a narrow-band absorption peak with an absorption rate exceeding 94% was obtained, and the device can be used as a biosensor with a sensitivity of 363 GHz/RIU. The designed THz metamaterials have broad application prospects in modulation, sensing, radar communication, and other fields.

A tunable optical buffer composed of a side-coupled racetrack microring resonator (RMRR) sequence and phase change material vanadium dioxide (VO₂) films has been designed and simulated with pass-through and drop functions. By optically induced phase transition of VO₂, the number of coupled RMRRs is dynamically controlled to realize the required delay time and the ultra-fast response. Simulation results show that the buffer with 12 RMRRs achieves a tunable delay of up to 309.6 ps, a working bandwidth of 306 GHz, and a super-fast response of 0.62 ps. The delay loss and power consumption are 0.012 dB / ps and 0.194 mW / ps, respectively. These findings suggest a new approach to realizing tunable delay in all-optical networks.

It was shown that the concentration of charge carriers and the stoichiometry of near-surface region are changed in p-type HgCdTe films with non-passivated and soiled by indium surface under storage temperatures 80°C and 120°C. The concentration of holes increased up to 10²⁴ cm ^{− 3} and mobility dropped up to 12 cm² / ( V × s ) under heating. From acquired Auger spectra, it was obtained that the mercury concentration decreased but the cadmium concentration increased in near-surface region as the indium spot approached. The surface passivation by photoresist prevented these degradation processes. The current–voltage characteristics of p-n junctions around the indium spots on the surface of p-type HgCdTe films passivated by SiO₂ / Si₃N₄ layers were measured after storage temperatures from 60°C up to 120° C. It was found that the p-n junctions closest to the indium spots have an increased dark current. The usage of an additional passivation layer of anodic oxide suppressed the appearance of regions with increased dark currents. It was concluded that the additional passivation layer of anodic oxide is necessary for durable and stable work of infrared detector arrays based on HgCdTe films.

The current field of autonomous driving has achieved superior object detection performance in good weather conditions. However, the environment sensing capability of autonomous vehicles is severely affected in rainfall traffic environments. Although deep-learning-based image derain algorithms have made significant progress, integrating them with high-level vision tasks, such as object detection, remains challenging due to the significant differences between the derain and object detection algorithms. Additionally, the accuracy of object detection in real rain traffic environments is significantly reduced due to the domain transfer problem between the training dataset and the actual rain environment. To address this domain-shifting problem, we propose an adaptive rain image enhancement object detection network for autonomous driving in adverse weather conditions (ARODNet). This network architecture consists of an image adaptive enhancement module, an image derain module, and an object detection module. The baseline detection module (CBAM-YOLOv7) is built by incorporating the YOLOv7 object detection network into a feed-forward convolutional neural network, and it includes an attention module (CBAM). We propose a domain adaptive rain image enhancement module, DRIP, for low-quality images acquired under heavy rainfall conditions. DRIP enhances low-quality images on rainy days by adaptively learning multiple preprocessing weights. To remove the effects of rain patterns and fog clouds on image detection, we introduce DRIP-enhanced images into the depth estimation derain module (DeRain) to prevent rain and fog from obscuring the objects to be detected. Finally, the multistage joint training strategy is adopted to improve the training efficiency, and the object detection is performed while the image is derained. The efficacy of the ARODNet network for object detection in rainy weather traffic environments has been demonstrated through numerous quantitative and qualitative studies.

Coherent optical communication critically relies on efficient digital signal processing (DSP). We demonstrate a lite DSP scheme to reduce the transmission cost and power consumption, which is mainly realized by reducing the sampling rate and simplifying the DSP algorithm. On the one hand, the baud-rate sampling technique based on the integral circuit can reduce the amount of data processing in DSP. On the other hand, we proposed a simplified joint modified constant modulus algorithm (MCMA) and phase-dependent decision-directed least mean square (DD-LMS) algorithm for adaptive blind polarization demultiplexing and phase recovery in which the butterfly structure is no longer needed for DD-LMS. And we compare the proposed simplified algorithm with the joint traditional CMA and blind phase search and the MCMA, in terms of performance in the 80-Gbaud dual-polarization 64QAM homodyne detection system. The results show that the proposed lite DSP mechanism can effectively reduce the power consumption of DSP by sub-rate sampling and simplifying the corresponding algorithm, which provides an alternative scheme for low-power optical interconnection of data centers.

To evaluate the effect of crosstalk caused by wavelength deviation of reconstruction-equivalent-chirp (REC) integrated multi-wavelength laser arrays (MWLAs) on an optical free-space communication system in which wavelength division multiplexing (WDM) is applied, we derive the model of power penalty caused by the wavelength deviation of REC-MWLAs and the bit error rate (BER) of the system under atmospheric turbulence and conduct a numerical simulation. The two adjacent channels in arranged REC laser arrays are defined as the analysis scope. We find that the power penalty is more sensitive to the deviation of the channel itself than the adjacent channel as the frequency shift increases from 0 to 50 GHz. The numerical relationship between the BER and power penalty is drawn to further investigate the tolerance for wavelength deviation under the requirements of BER < 10 ^{− 6} and BER < ^{10 − 9}. In addition, for a smaller deviation of 5 GHz of the analyzed channel, the linewidth compression of the REC laser can decrease the power penalty from 1.3 to 0.7 dB, improving the system performance and moderating the demands for deviation of REC-MWLAs, and the gemination of channel spacing does not evidently reduce the requirement for deviation. The results of this work demonstrate the feasibility of REC-MWLAs as the light source of WDM free-space communication systems under atmospheric turbulence and provide an effective reference for the allowable deviation in the design and testing of MWLAs.

We propose an adaptive localization and channel state information (CSI) estimation model using the distance vector for indoor multiple-input multiple-output (MIMO) visible light communication (VLC). Under normal conditions, receivers (Rxs) require the CSI of all channels [formed by multiple transmitters (Txs) and multiple Rxs] to recover the original data, and the received signal strength (RSS) obtained from Txs, i.e., light-emitting diodes (LEDs) to predict the Rx’s location. A beacon [or pilot signal (PS)] with CSI is periodically broadcasted from each LED to MIMO Rxs for constructing the CSI matrix and measuring the RSS. However, in an abnormal condition, PSs from one or multiple Txs might not reach Rxs due to shadowing, thus resulting in a failure in positioning and data recovery. To combat this, the proposed model enables the Rx to predict its location without the need for all PSs based on the construction of the distance vectors at Rx. Simulations conducted for two scenarios of user’s mobility, i.e., lattice and random direction mobility, show a low positioning error of ∼0.1 and 0.2 m, respectively. Based on predicted positions, MIMO VLC is capable of fulfilling the CSI matrix and assists the Rx in recovering transmitted data with a very low bit error rate.

Compared with visible light imaging technology, infrared optical imaging technology (see Fig. 1) is less affected by weather and illumination and is a potential auxiliary solution method for the future of the autonomous driving system. Based on the above characteristics, this study proposes the hold surrounding’s key (HSK)-you only look once (YOLOv7) algorithm. First, based on the characteristics of the infrared image, this study optimizes the network structure of YOLOv7 and proposes a vehicle and pedestrian detection algorithm based on the improved YOLOv7. Aiming at the problem that the reasoning speed and detection accuracy of pedestrian and vehicle detection algorithms based on infrared images are challenging to balance, occupy large storage space, and are difficult to deploy and run in real-time in low and medium-performance devices, the MPConv is added to replace the Conv structure in YOLOv7. In view of the false detection, missed detection, mutual occlusion and overlap of detected objects, and other situations that the YOLOv7 algorithm is prone to cause in the actual deployment environment easily, a tiny object detection layer is added. At the same time, to solve the problem that infrared optical imaging systems are prone to noise caused by external factors, this study introduces the TRPCA method for image denoising in the preprocessing process of the YOLOv7 algorithm. In the end, the HSK-YOLOv7 algorithm is verified using the self-made infrared traffic object detection dataset and the publicly available FLIR dataset to verify the detection effect of the HSK-YOLOv7 algorithm on near-infrared images and thermal infrared images. The parameter quantity of our algorithm is 37.3M, and the computing throughput is 107.5 GFLOPs. The detection speed on the self-made dataset and FLIR dataset reaches 163 frames per second (FPS) and 71.8 FPS, respectively, and the mAP@0.5 indicator reaches 94.08% and 61.3%, respectively. In general, HSK-YOLOv7 can meet the real-time requirements of the autonomous driving system while ensuring detection accuracy.

Erratum corrects an error in Fig. 8.