SSD Algorithm Research Articles

A method of fingerprint indoor localization based on received signal strength difference by using compressive sensing

With the development of network technology, WLAN-based indoor localization plays an increasingly important role. Most current localization methods are based on the comparison between the received signal strength indication (RSSI) and the RSS in the database, whose nearest reference point is the location point. However, since a uniform standard for measuring components of smartphones has not yet been established, the Wi-Fi chipsets on different smartphones may have different sensitivity levels to different Wi-Fi access points (APs) and channels. Even for the same signal, RSSI values obtained by different terminals at the same time and the same location may be different. Therefore, the impact of terminal heterogeneity on localization accuracy can be overlooked. To address this issue, a fusion method based on received signal strength difference and compressive sensing (RSSD-CS) is proposed in this paper, which can reduce the influence caused by the terminal heterogeneity. Besides, a fingerprint database is reconstructed from the existing reference point data. Experiments show that the proposed RSSD-CS algorithm can achieve high localization accuracy in indoor localization, and the accuracy is enhanced by 20.5% and 15.6% compared to SSD and CS algorithm.

Target part detection based on improved SSD algorithm

Aiming at the problem that the traditional target detection algorithm cannot balance the target detection accuracy and real-time detection, and the detection effect is poor under the actual complex production conditions. A deep learning detection method based on Inception-SSD framework is proposed. The Inception-SSD, which integrates the Inception prediction structure, introduces the Inception block into the extra layer of the SSD network. Then it uses batch normalization (BN) and residual structure connection to capture more target information without increasing the complexity of the network. It will improve its accuracy without affecting its real-time detection. The result of the experiment show that the detection accuracy of this model is 97.8% in the actual production process, which is largely improved comparing with the original SSD algorithm. The detection time is 48ms, which improves the detection accuracy and ensures the real-time performance. It can meet the detection requirements of parts in actual production.

Improving the prediction rate of unusual behaviors of animal in a poultry using deep learning technique

Poultry farms across the world house animals such as cows, sheep’s, pigs and hen These farms are the places from where the market gets meat, eggs, wool and other animal products which are used in our daily lives. But sometimes the business of the farms is severely hit primarily due to poultry animals catching diseases or these animals succumbing to injuries caused due to fighting each other. The identification of such animals showing unusual characteristics or behavior is necessary. We through our paper use deep learning concepts to help the poultry owner to identify these unusual characteristics in a sheep. Some of the characteristics that our paper identifies are skinny, redness, non-aggression, aggression, beefy and foraging. Our paper makes use of a video, shot from a camera and the implementation of a sequential model as well as make use of the SSD algorithm to identify and characterize the sheep’s depicted in the video. The sequential model is trained using the training dataset which contains static images of sheep’s and the attribute/characteristics depicted by those sheep’s that are stored in a CSV file. The testing dataset contains the images that are extracted from the input video as frames. The testing dataset is passed through the sequential model to get the characteristics/attributes depicted by the sheep in each frame and store those characteristics/attributes in a CSV file. The SSD algorithm is trained on identifying the various animals and not only does it display the name of the animal detected, it displays the confidence percentage of the animal as well, in our paper it is used for identifying sheep. The above stated algorithm also creates a border around the identified sheep such that it can be used for tracking purposes during the entirety of the video. The SSD algorithm also takes in the attributes depicted by these frames and displays them along the border which identifies the sheep. The accuracy is compared with YOLO algorithms and shows 3 to 6% improvement on prediction rate.

Research on Detection Technology of Autonomous Landing Based on Airborne Vision

In this paper, aiming at the cooperative target detection problem in the process of unmanned helicopter sliding down, a detection method based on complementary filtering is proposed, which fuses improved SSD algorithm and related filtering KCF algorithm. The improved deep learning SSD model redesigns the feature extraction structure to improve the detection effect of small and medium targets for the small target size and large scale change in the landing scene. Then use the detection results of the SSD model to correct the KCF detection, adjust the weight parameters, and output the final fusion detection results. The test results show that the improved model detection accuracy is significantly improved, the detection accuracy in various environments reaches 93.3%, which is higher than 86.1% of the classic SSD model and 87.5% of the Faster-rcnn model. The final proposed fusion detection algorithm has a success rate of 91.1% and a processing speed of 91 hz, which basically satisfies the requirements of the ship.

Vehicle and wheel detection: a novel SSD-based approach and associated large-scale benchmark dataset

Many vehicle and wheel detection methods based on convolutional neural networks suffer from challenges due to the lack of training data and the limitation on small object detection. To solve this problem, we present a novel optimized SSD algorithm with multi-concatenation modules, aiming to improve the performance of small object detection. In the multi-concatenation module, features from different layers are concatenated together, including feature map from shallow layer with more location information, feature map from intermediate layer, and feature map from deep layer with rich semantic information. SEBlock is employed to re-weight the new feature map to improve the quality of representation. Furthermore, to facilitate the study of vision-based vehicle and wheel detection, a large-scale benchmark dataset of 8209 images is established, comprising five object categories: truck, pickup, tractor, car, and wheel. On the Pascal VOC 2007 test set, our network achieves 78.7% mAP, which is higher than SSD by 1.5%. On KITTI dataset, the proposed method can reach 71.4% mAP, surpassing SSD by 3.5%. In addition, experimental results show that the proposed method results in better detection performance on small objects.

A deep model method for recognizing activities of workers on offshore drilling platform by multistage convolutional pose machine

The growing diversity of image scenes brings a great challenge to human activity recognition in practice. Traditional activity recognition methods cannot satisfy the demand of precise action recognition in complex scenes. In this work, we build a training set of worker's activities on offshore drilling platform by collecting data from offshore drilling monitor, and then an improved multi-level convolutional pose machine (MCPM) method is proposed and trained to recognize activities of workers on the platforms. In human object detection, a multi-rule region proposal marker algorithm is developed to separate the seawater area, and the ducts of similar personnel is pre-discriminated by support vector machine. We use the characteristics of the human body key-points not affected by complex background noise to assist the detection of the human target. As results, it shown that our method performs better than Faster-RCNN, MobileNet-SSD and SSD algorithms in detecting human target on the offshore drilling platform, and achieves well accuracy in recognizing many key activities. To our best acknowledge, it is the first attempt of using deep model to recognize worker's activities on offshore drilling platform.

Multi-Feature Fusion and Enhancement Single Shot Detector for Traffic Sign Recognition

Road traffic sign detection and recognition play an important role in advanced driver assistance systems (ADAS) by providing real-time road sign perception information. In this paper, we propose an improved (Single Shot Detector) SSD algorithm via multi-feature fusion and enhancement, named MF-SSD, for traffic sign recognition. First, low-level features are fused into high-level features to improve the detection performance of small targets in the SSD. We then enhance the features in different channels to detect the target by enhancing effective channel features and suppressing invalid channel features. Our algorithm gets good results in domestic real-time traffic signs. The proposed MF-SSD algorithm is evaluated with the German Traffic Sign Recognition Benchmark (GTSRB) dataset. The experimental results show that the MF-SSD algorithm has advantages in detecting small traffic signs. Compared with existing methods, it achieves higher detection accuracy, better efficiency, and better robustness in complex traffic environment.

Design of Smart Unstaffed Retail Shop Based on IoT and Artificial Intelligence

Unstaffed retail shops have emerged recently and been noticeably changing our shopping styles. In terms of these shops, the design of vending machine is critical to user shopping experience. The conventional design typically uses weighing sensors incapable of sensing what the customer is taking. In the present study, a smart unstaffed retail shop scheme is proposed based on artificial intelligence and the internet of things, as an attempt to enhance the user shopping experience remarkably. To analyze multiple target features of commodities, the SSD ( $300\times 300$ ) algorithm is employed; the recognition accuracy is further enhanced by adding sub-prediction structure. Using the data set of 18, 000 images in different practical scenarios containing 20 different type of stock keeping units, the comparison experimental results reveal that the proposed SSD ( $300\times 300$ ) model outperforms than the original SSD ( $300\times 300$ ) in goods detection, the mean average precision of the developed method reaches 96.1% on the test dataset, revealing that the system can make up for the deficiency of conventional unmanned container. The practical test shows that the system can meet the requirements of new retail, which greatly increases the customer flow and transaction volume.

特征增强SSD算法及其在遥感目标检测中的应用

特征增强SSD算法及其在遥感目标检测中的应用

Pipeline Magnetic Flux Leakage Image Detection Algorithm Based on Multiscale SSD Network

In order to solve the problem of low detection accuracy of small targets in the SSD detection algorithm, a pipeline magnetic flux leakage image detection algorithm based on multiscale SSD network is proposed in this paper. The dilated convolution and attention residual module are introduced into the SSD algorithm to fuse the low-resolution high-semantic information feature map with the high-resolution low-semantic information feature map so as to improve the resolution of the low-resolution feature map and provide detailed features for small targets. Finally, the target location and category are obtained by regression algorithm. The experimental results show that the proposed algorithm can automatically identify the location of circumferential weld, spiral weld, and defect of magnetic flux leakage data. Compared with the original SSD algorithm framework, the improved algorithm has higher detection accuracy, 97.62%, 18.01% lower false detection rate, 18.36% lower false detection rate, better robustness, and obvious effect on small target detection.

Object Detection Based on the Improved Single Shot MultiBox Detector

Aiming at the poor effect of deep learning algorithm on small objects detection, the SSD object detection method based on feature fusion is proposed. The reasons for low detection rate and poor robustness of classical SSD object detection methods are analysed; and through the theoretical analysis and comparative experiments, the characteristic fusion layer was proposed. The shallow layers with high resolution and deep layers with strong semantics are fused with the feature fusion structure; finally, a complete feature fusion structure is designed with the residual block to increase the width and depth of the network. The contrast experiment on the PASCAL VOC dataset was conducted for detection capability and detection accuracy, and experimental result indicates that when the confidence is set to 0.5, the mAP of the SSD method based on feature fusion is 78.04%, which is 0.8% higher than the classical SSD algorithm and 4.8% higher than the Faster RCNN algorithm. Obviously, the proposed algorithm improves the ability of small objects, and verifies the effectiveness of the proposed algorithm.

改进的SSD算法及其对遥感影像小目标检测性能的分析

改进的SSD算法及其对遥感影像小目标检测性能的分析

Research on Simulation System of Small Unmanned Air-to-Ground Object Detection Platform

Air-to-ground object detection has important application value in many fields such as military reconnaissance, air strike, personnel rescue, and natural exploration. Object detection algorithm based on deep learning solves the problems of poor generality and low robustness of traditional manual detection algorithms. It has great advantages in air-to-ground detection tasks, but for small unmanned platforms, currently deep learning object detection algorithms have high hardware requirements and are difficult to deploy. In this paper, a deep learning object detection simulation system is built based on the TX2 embedded development board with the features of low power consumption, light weight and fast calculation speed. This object detection simulation experiment is performed for two mainstream deep learning object detection algorithms. The experimental results showed that the regression-based SSD algorithm combined with MobileNets feature extraction network could perform real-time in the embedded system, which laid the foundation for the actual construction of the next unmanned aerial platform.

基于改进SSD的轻量化小目标检测算法

In order to improve the small object detection ability of SSD object detection algorithm, the transposed convolution structure in SSD algorithm was proposed, the low resolution high semantic information feature map was integrated with high resolution low semantic information feature map using transposed convolution, which increased the ability of low level feature extraction and improved the average accuracy of SSD algorithm. At the same time for the problem that SSD algorithm model being large, running memory consumption high, without running on the embedded equipment ARM, a lightweight feature extraction minimum unit was proposed based on DenseNet, combining depthwise separable convolutions, pointwise group convolution and channel shuffle, running on the embedded equipment ARM cloud be realized. The comparative experiments on PASCAL VOC data set and KITTI autopilot data set show that the mean average is significantly improved by improved network structure, and the number of model parameters is effectively reduced.

Calculation Principle and Test Applications for Digital Speckle Correlation Technology

With SSD algorithm and calculus process in detail computing, the basic calculation principles and the solution process from digital image to displacement of digital speckle correlation technology were described. Then the digital speckle correlation technology were used to the tensile test of CFRP plate embedded reinforcement beams, obtained the displacement cloud image and destruction rules of the specimens. The results show that the calculation results analyzed by the digital image correlation technology are in good agreement with the experimental phenomena. And the technology has the characteristic of full field measurement, simple operation, strong adaptability and high precision, so it has the widespread application prospect, especially suitable for application in the mechanical properties of materials under complex conditions, a large section of the destruction process monitoring and other fields.

Dynamic scene에 대한 카메라 정보 추출 기법

본 논문에서는 동적 영상에서 이동하는 관심 물체의 위치좌표를 추출하고 영역화할 수 있는 자동 3D 촬영 시스템의 카메라 정보 추출 기법을 제안하였다. 즉, 기존 블록 기반의 MSE (mean square error) 알고리즘을 이용하여 초기화된 기준영상과 현재 입력되는 스테레오 영상간의 비교를 통해 좌, 우측 관심 물체의 위치좌표를 추출한 다음, 이들 추출된 위치좌표를 이용하여 스테레오 카메라의 팬/틸트를 제어하고, 제어된 팬/틸트의 움직임 각도와 스테레오 카메라의 기하학적인 구성요소를 이용하여 관심물체의 마스크 크기를 결정함으로써 다음 프레임의 기준영상으로 사용될 표적 영역화가 이루어진다. 새롭게 영역화된 기준영상은 이후 같은 방법으로 갱신되며, 검출된 위치좌표에 따라 스테레오 카메라의 주시각 제어 및 FOV 제어를 통해 실시간 자동 추적이 이루어지게 된다.

Extracting Depth Information Using a Correlation Matching Algorithm

This manuscript presents a modified algorithm to extract depth information from stereo-vision acquisitions using a correlation based approaches. The main implementation of the proposed method is in the area of autonomous Pick & Place, using a robotic manipulator. Current vision-guided robotics is still based on a priori training and teaching steps, and still suffers from long response time. This study uses a stereo triangulation setup where two Charged Coupled Devices CCDs are arranged to acquire the scene from two different perspectives. The study discusses the details to calculate the depth using a correlation matching routine which programmed using a Square Sum Difference SSD algorithm to search for the corresponding points from the left and the right images. The SSD is further modified using an adjustable Region Of Interest ROI along with a center of gravity based calculations. Furthermore, the two perspective images are rectified to reduce the required processing time. The reported error in depth using the modified SSD method is found to be around 1.2 mm.

A fast and robust feature-based 3D algorithm using compressed image correlation

Two objectives of 3D computer vision are high processing speed and precise recovery of object boundaries. This paper addresses these issues by presenting an algorithm that combines feature-based 3D matching with Compressed Image Correlation. The algorithm uses an image compression scheme that retains pixel values in high intensity gradient areas while eliminating pixels with little correlation information in smooth surface regions. The remaining pixels are stored in sparse format along with their relative locations encoded into 32-bit words. The result is a highly reduced image data set containing distinct features at object boundaries. Consequently, far fewer memory calls and data entry comparisons are required to accurately determine edge movement. In addition, by utilizing an error correlation function, pixel comparisons are made through single integer calculations eliminating time consuming multiplication and floating point arithmetic. Thus, this algorithm typically results in much higher correlation speeds than spectral correlation and SSD algorithms. Unlike the traditional fixed window sorting scheme, adaptive correlation window positioning is implemented by dynamically placing object boundaries at the center of each correlation window. Processing speed is further improved by compressing and correlating the images in only the direction of disparity motion between frames. Test results on both simulated disparities and real motion image pair are presented.

Pelviscanner hélicoïdal en basse dose : évaluation de la dose d’irradiation et du traitement d’images

Purpose To estimate from phantom measurements the radiation dose and the accuracy of helical CT pelvimetry. Materials and methods Eight helical CT acquisitions using different tube current (100, 50, 25 or 10mAs) and pitch factor (1.125 or 1.375) settings but identical collimation (2mm) and kilovoltage (120 kVp) were evaluated using a four-channel MDCT scanner and compared with conventional CT pelvimetry including a single scout and two transverse images. A plexiglas phantom combined with an ionization chamber was used to calculate the CTDIw and DLP for each acquisition. Then, an ex vivo phantom of bony pelvis was used to evaluate the accuracy of helical acquisitions for the measurement of pelvic diameters (i.e. the antero-posterior inlet, the transverse inlet and the interspinous distance). Reconstructions of helical acquisitions were performed using 2D MPR, 3D MIP and 3D SSD algorithms. Results CTDIw and DLP of conventional pelvimetry were 26 mGy and 42 mGy.cm respectively. The radiation dose of helical acquisitions decreased linearly with tube current (CTDIw: from 13 to 1.3 mGy, DLP: from 218.3 to 18.7 mGy.cm). Compared to conventional CT, the dose was nearly similar at 25 mAs and reduced at 10 mAs. Helical acquisitions provided accurate measurements of pelvic diameters with a pitch of 1.125 and a 2D MPR algorithm to evaluate the AP inlet and a 3D MIP algorithm to evaluate the transverse inlet and the interspinous distance. Variations of tube current did not influence the accuracy of pelvic diameter measurement. Conclusion Our results suggest that accurate low-dose helical CT pelvimetry using 10-25 mAs and a pitch factor of 1.125 combined with 2D MPR and 3D MIP reconstructions is possible.

A simple equivalent tissue-air ratio method for calculating absorbed dose in a heterogeneous medium.

Density correction factors for 60Co gamma rays and 8-MV x rays were measured with an ionization chamber in water-equivalent phantoms containing simulated lung or bone inhomogeneities larger or smaller than the beam (beam sizes, 5 x 5 to 30 x 30 cm). Measurements were compared with values calculated from the effective SSD method, the generalized Batho power law method, and a simplified equivalent tissue-air ratio (SETAR) method. The results indicate that the SETAR method gives the best overall agreement with experimental values, while the effective SSD and Batho algorithms are accurate only in certain situations. The SETAR method seems adequate for most clinical applications.