INRIA Dataset Research Articles

A Residual-Inception U-Net (RIU-Net) Approach and Comparisons with U-Shaped CNN and Transformer Models for Building Segmentation from High-Resolution Satellite Images.

Building segmentation is crucial for applications extending from map production to urban planning. Nowadays, it is still a challenge due to CNNs’ inability to model global context and Transformers’ high memory need. In this study, 10 CNN and Transformer models were generated, and comparisons were realized. Alongside our proposed Residual-Inception U-Net (RIU-Net), U-Net, Residual U-Net, and Attention Residual U-Net, four CNN architectures (Inception, Inception-ResNet, Xception, and MobileNet) were implemented as encoders to U-Net-based models. Lastly, two Transformer-based approaches (Trans U-Net and Swin U-Net) were also used. Massachusetts Buildings Dataset and Inria Aerial Image Labeling Dataset were used for training and evaluation. On Inria dataset, RIU-Net achieved the highest IoU score, F1 score, and test accuracy, with 0.6736, 0.7868, and 92.23%, respectively. On Massachusetts Small dataset, Attention Residual U-Net achieved the highest IoU and F1 scores, with 0.6218 and 0.7606, and Trans U-Net reached the highest test accuracy, with 94.26%. On Massachusetts Large dataset, Residual U-Net accomplished the highest IoU and F1 scores, with 0.6165 and 0.7565, and Attention Residual U-Net attained the highest test accuracy, with 93.81%. The results showed that RIU-Net was significantly successful on Inria dataset. On Massachusetts datasets, Residual U-Net, Attention Residual U-Net, and Trans U-Net provided successful results.

CrossGeoNet: A Framework for Building Footprint Generation of Label-Scarce Geographical Regions

• Planet satellite imagery is utilized for building mapping in less developed regions (e.g., cities in Africa) • A co-segmentation learning framework is proposed to tackle the problem of insufficient labeled samples in target cities for the task of building footprint generation • A cross-geolocation attention module is proposed to effectively learn the underlying similarity between different geolocations Building footprints are essential for understanding urban dynamics. Planet satellite imagery with daily repetition frequency and high resolution has opened new opportunities for building mapping at large scales. However, suitable building mapping methods are scarce for less developed regions, as these regions lack massive annotated samples to provide strong supervisory information. To address this problem, we propose to learn cross-geolocation attention maps in a co-segmentation network, which is able to improve the discriminability of buildings within the target city and provide a more general building representation in different cities. In this way, the limited supervisory information resulting from insufficient training examples in target cities can be compensated. Our method is termed as CrossGeoNet, and consists of three elemental modules: a Siamese encoder, a cross-geolocation attention module, and a Siamese decoder. More specifically, the encoder learns feature maps from a pair of images from two different geo-locations. The cross-location attention module aims at learning similarity based on these two feature maps and can provide a global overview of common objects (e.g., buildings) in different cities. The decoder predicts segmentation masks of buildings using the learned cross-location attention maps and the original convolved images. The proposed method is evaluated on two datasets with different spatial resolutions, i.e., Planet dataset (3 m/pixel) and Inria dataset (0.3 m/pixel), which are collected from various locations around the world. Experimental results show that CrossGeoNet can well extract buildings of different sizes and alleviate false detections, which significantly outperforms other competitors.

An Intelligent System for Detecting Abnormal Behavior in Students Based on the Human Skeleton and Deep Learning.

With the use of an intelligent video system, this research provides a method for detecting abnormal behavior based on the human skeleton and deep learning. To begin with, the spatiotemporal features of human bones are extracted through iterative training using the OpenPose deep learning network and the redundant information of human bone facial features is reduced in the feature extraction process, effectively reducing the time it takes to identify and analyze abnormal behavior. The collected human skeleton features are then classified using a graph convolution neural network to reduce the computational complexity of the behavior identification algorithm, and the sliding window voting method is used to further improve the accuracy of the behavior classification in practical application, resulting in the diagnosis and classification of abnormal behavior of students under video surveillance. Finally, using the self-built student trajectory data set and the INRIA data set, simulation analysis is performed, and the practicality and superiority of the proposed method for abnormal behavior detection is confirmed by comparing it to the existing abnormal behavior recognition methods. The proposed method for detecting anomalous behavior in a self-built database and INRIA data set has a high accuracy of more than 99.50 percent and a high processing efficiency rate.

Comparison of residual and dense neural network approaches for building extraction from high-resolution aerial images

Applications including change detection, disaster management, and urban planning require precise building information, and therefore automatic building extraction has become a significant research topic. With the improvements in sensor and satellite technologies, more data has become available, and with the increased computational power, deep learning methods have emerged as successful tools. In this study, U-Net and FPN architectures using four different backbones (ResNet-50, ResNeXt-50, SE-ResNext-50, and DenseNet-121), and an Attention Residual U-Net approach were used for building extraction from high-resolution aerial images. Two publicly available datasets, Inria Aerial Image Labeling Dataset and Massachusetts Buildings Dataset were used to train and test the models. According to the results, Attention Residual U-Net model has the highest F1 score with 0.8154, IoU score with 0.7102, and test accuracy with 94.51% on the Inria dataset. On the Massachusetts dataset, FPN Dense-Net-121 model has the highest F1 score with 0.7565 and IoU score with 0.6188, and Attention Residual U-Net model has the highest test accuracy with 92.43%. It has been observed that, FPN with DenseNet backbone can be a better choice when working with small size datasets. On the other hand, Attention Residual U-Net approach achieved higher success when a sufficiently large dataset is provided.

Comparative research on different backbone architectures of DeepLabV3+ for building segmentation

Rapid urban growth and globalization affect land use in cities, and the need for automatic interpretation of remote sensing images is constantly increasing. Deep neural networks are becoming widespread in high-resolution aerial and satellite image sources in Earth observation missions. Various convolutional neural network (CNN) architectures have been implemented in building extraction, but it is still challenging to distinguish building class from other man-made classes in public datasets. Here, we present comparative research for automatic building extraction on different data sources using DeepLabV3+ architecture with ResNet-18, ResNet-50, Xception, and MobileNetv2 models. The CNNs are implemented on Inria Aerial Image Labeling, Massachusetts Buildings, and Wuhan University Building Extraction Datasets in terms of evaluation metrics and training and testing time consumption. Our implementation of the DeepLabV3 + ResNet-50 model performed F1-score of 97.44% in Massachusetts Building dataset and intersection over union as 80.75% in Inria dataset, higher than at least 3% than the previous studies.

RegGAN: An End-to-End Network for Building Footprint Generation with Boundary Regularization

Accurate and reliable building footprint maps are of great interest in many applications, e.g., urban monitoring, 3D building modeling, and geographical database updating. When compared to traditional methods, the deep-learning-based semantic segmentation networks have largely boosted the performance of building footprint generation. However, they still are not capable of delineating structured building footprints. Most existing studies dealing with this issue are based on two steps, which regularize building boundaries after the semantic segmentation networks are implemented, making the whole pipeline inefficient. To address this, we propose an end-to-end network for the building footprint generation with boundary regularization, which is termed RegGAN. Our method is based on a generative adversarial network (GAN). Specifically, a multiscale discriminator is proposed to distinguish the input between false and true, and a generator is utilized to learn from the discriminator’s response to generate more realistic building footprints. We propose to incorporate regularized loss in the objective function of RegGAN, in order to further enhance sharp building boundaries. The proposed method is evaluated on two datasets with varying spatial resolutions: the INRIA dataset (30 cm/pixel) and the ISPRS dataset (5 cm/pixel). Experimental results show that RegGAN is able to well preserve regular shapes and sharp building boundaries, which outperforms other competitors.

MS-ML-SNYOLOv3: A robust lightweight modification of SqueezeNet based YOLOv3 for pedestrian detection

Pedestrian detection is a vital problem area in computer vision applications such as surveillance, automotive safety, and robotics. To continue the progress made in this field, a deep learning based approach is given in this work. A modified lightweight YOLOv3 is proposed, which is named as MultiScale-MultiLevel-SqueezeNetYOLOv3 (MS-ML-SNYOLOv3). An improved base network forms the basis of the proposed YOLOv3. Two Squeeze & Expand blocks are embodied in SqueezeNet at specific levels of the network to extract a hierarchical feature representation. An addition of a larger receptive field is made in the Expand section as well to encapsulate the feature information at various scales. The proposed method is compared against state-of-the-art detection methods, Faster RCNN, YOLOv2, YOLOv3, and SSD. The comparison also includes various base networks for YOLOv2 and YOLOv3. The performance metrics used for comparison are Detection Error Trade-Off, and Precision-Recall Curve. The proposed method is evaluated on benchmark datasets INRIA and Caltech Pedestrian datasets. A statistical analysis, Friedman Test is conducted to verify the application of MS-ML-SNYOLOv3 in real world data.

PRNet++: Learning towards generalized occluded pedestrian detection via progressive refinement network

Pedestrian detection has achieved significant progress in recent years. Though promising results have been obtained on standard pedestrians, it remains challenging to detect pedestrians in various occlusion situations. In this paper, we propose Progressive Refinement Network (PRNet), a novel single-stage detector for occluded pedestrian detection. Inspired by human’s progressive process on annotating occluded pedestrians, PRNet perform sequential refinement in a single-stage detection framework by three phases: finding anchors of visible parts with high confidence, calibrating these anchors with a full body template derived from occlusion statistics, and adjusting the calibrated anchors to target full-body regions. Unlike conventional methods that utilize predefined anchors directly for full-body estimation, the proposed confidence-aware anchor calibration offers an adaptive anchor initialization for detection with occlusions, while helps reduce the gap between visible-part and full-body detection. We also propose an occlusion loss, which automatically up-weights heavily occluded pedestrian samples. In addition, a Receptive Field Backfeed (RFB) module is introduced to diversify receptive fields in early layers that commonly fire only on visible parts or small-size full-body regions. To further learn generalized representations of pedestrians in different occlusion states, we propose to establish a new single-stage detector with dual-stream architecture namely PRNet++. An Easy-branch and a Hard-branch are designed to learn complementary representation that are more robust to various occlusions. Moreover, the generalization ability of models for other domains are critical in real-world applications, but it has not attracted too much attention. To address this issue, we introduce the unsupervised domain adaptation setting. To handle occlusion situations in unknown domain better, we especially design a Dynamic Iterative adaptation strategy and a Multi-experts adaptation strategy for our occlusion-aware detectors PRNet and PRNet++. Extensive supervised within-dataset experiments and unsupervised domain adaptation experiments were performed on CityPersons, Caltech, ECP-night, KITTI and INRIA datasets, which can validate the effectiveness of our proposed methods in occluded pedestrian detection.

Building extraction from remote sensing images using deep residual U-Net

ABSTRACT Building extraction is a fundamental area of research in the field of remote sensing. In this paper, we propose an efficient model called residual U-Net (RU-Net) to extract buildings. It combines the advantages of U-Net, residual learning, atrous spatial pyramid pooling, and focal loss. The U-Net model, based on modified residual learning, can reduce the parameters and degradation of the network; atrous spatial pyramid pooling can acquire multiscale features and context information of the sensing images; and focal loss can solve the problem of unbalanced categories in classification. We implemented it on the WHU aerial image dataset and the Inria aerial image labeling dataset. The results of RU-Net were compared with the results of U-Net, FastFCN, DeepLabV3+, Web-Net, and SegNet. Experimental results show that the proposed RU-Net is superior to the others in all aspects of the WHU dataset. For the Inria dataset, most evaluation metrics of RU-Net are better than the others, as well as the sharp, boundary, and multiscale information. Compared with FastFCN and DeepLabV3+, our method increases the efficiency by three to four times.

Semi-Supervised Building Footprint Generation With Feature and Output Consistency Training

Accurate and reliable building footprint maps are vital to urban planning and monitoring, and most existing approaches fall back on convolutional neural networks (CNNs) for building footprint generation. However, one limitation of these methods is that they require strong supervisory information from massive annotated samples for network learning. State-of-the-art semi-supervised semantic segmentation networks with consistency training can help to deal with this issue by leveraging a large amount of unlabeled data, which encourages the consistency of model output on data perturbation. Considering that rich information is also encoded in feature maps, we propose to integrate the consistency of both features and outputs in the end-to-end network training of unlabeled samples, enabling to impose additional constraints. Prior semi-supervised semantic segmentation networks have established the cluster assumption, in which the decision boundary should lie in the vicinity of low sample density. In this work, we observe that for building footprint generation, the low-density regions are more apparent at the intermediate feature representations within the encoder than the encoder's input or output. Therefore, we propose an instruction to assign the perturbation to the intermediate feature representations within the encoder, which considers the spatial resolution of input remote sensing imagery and the mean size of individual buildings in the study area. The proposed method is evaluated on three datasets with different resolutions: Planet dataset (3 m/pixel), Massachusetts dataset (1 m/pixel), and Inria dataset (0.3 m/pixel). Experimental results show that the proposed approach can well extract more complete building structures and alleviate omission errors.

Building Footprint Generation Through Convolutional Neural Networks With Attraction Field Representation

Building footprint generation is a vital task in a wide range of applications, including, to name a few, land use management, urban planning and monitoring, and geographical database updating. Most existing approaches addressing this problem fall back on convolutional neural networks (CNNs) to learn semantic masks of buildings. However, one limitation of their results is blurred building boundaries. To address this, we propose to learn attraction field representation for building boundaries, which is capable of providing an enhanced representation power. Our method comprises two elemental modules: an Img2AFM module and an AFM2Mask module. More specifically, the former aims at learning an attraction field representation conditioned on an input image, which is capable of enhancing building boundaries and suppressing the background. The latter module predicts segmentation masks of buildings using the learned attraction field map. The proposed method is evaluated on three datasets with different spatial resolutions: the ISPRS dataset, the INRIA dataset, and the Planet dataset. From experimental results, we find that the proposed framework can well preserve geometric shapes and sharp boundaries of buildings, which brings significant improvements over other competitors. The trained model and code are available at https://github.com/lqycrystal/AFM_building.

Robust baggage detection and classification based on local tri-directional pattern

In recent decades, the automatic video surveillance system has gained significant importance in computer vision community. The crucial objective of surveillance is monitoring and security in public places. In the traditional Local Binary Pattern, the feature description is somehow inaccurate, and the feature size is large enough. Therefore, to overcome these shortcomings, our research proposed a detection algorithm for a human with or without carrying baggage. The Local tri-directional pattern descriptor is exhibited to extract features of different human body parts including head, trunk, and limbs. Then with the help of support vector machine, extracted features are trained and evaluated. Experimental results on INRIA and MSMT17 V1 datasets show that LtriDP outperforms several state-of-the-art feature descriptors and validate its effectiveness.

Pedestrian detection algorithm based on improved muti-scale feature fusion

Aiming at the problem of false detection and missed detection of small targets and occluded targets in the process of pedestrian detection, a pedestrian detection algorithm based on improved multi-scale feature fusion is proposed. First, for the YOLOv4 multi-scale feature fusion module PANet, which does not consider the interaction relationship between scales, PANet is improved to reduce the semantic gap between scales, and the attention mechanism is introduced to learn the importance of different layers to strengthen feature fusion; then, dilated convolution is introduced. Dilated convolution reduces the problem of information loss during the downsampling process; finally, the K-means clustering algorithm is used to redesign the anchor box and modify the loss function to detect a single category. The experimental results show that the improved pedestrian detection algorithm in the INRIA and WiderPerson data sets under different congestion conditions, the AP reaches 96.83% and 59.67%, respectively. Compared with the pedestrian detection results of the YOLOv4 model, the algorithm improves by 2.41% and 1.03%, respectively. The problem of false detection and missed detection of small targets and occlusion has been significantly improved.

A Shallow–Deep Feature Fusion Method for Pedestrian Detection

In this paper, a shallow–deep feature fusion (SDFF) method is developed for pedestrian detection. Firstly, we propose a shallow feature-based method under the ACF framework of pedestrian detection. More precisely, improved Haar-like templates with Local FDA learning are used to filter the channel maps of ACF such that these Haar-like features are able to improve the discriminative power and therefore enhance the detection performance. The proposed shallow feature is also referred to as weighted subset-haar-like feature. It is efficient in pedestrian detection with a high recall rate and precise localization. Secondly, the proposed shallow feature-based detection method operates as a region proposal. A classifier equipped with ResNet is then used to refine the region proposals to judge whether each region contains a pedestrian or not. The extensive experiments evaluated on INRIA, Caltech, and TUD-Brussel datasets show that SDFF is an effective and efficient method for pedestrian detection.

Semantic segmentation from remote sensor data and the exploitation of latent learning for classification of auxiliary tasks

In this paper we address three different aspects of semantic segmentation from remote sensor data using deep neural networks. Firstly, we focus on the semantic segmentation of buildings from remote sensor data and propose ICT-Net: a novel network with the underlying architecture of a fully convolutional network, infused with feature re-calibrated Dense blocks at each layer. Uniquely, the proposed network combines the localization accuracy and use of context of the U-Net network architecture, the compact internal representations and reduced feature redundancy of the Dense blocks, and the dynamic channel-wise feature re-weighting of the Squeeze-and-Excitation(SE) blocks. The proposed network has been tested on the INRIA and AIRS benchmark datasets and is shown to outperform other state of the art. Secondly, as the building classification is typically the first step of the reconstruction process, we investigate the relationship of the classification accuracy to the reconstruction accuracy. Due to the lack of (1) scene depth information, and (2) ground-truth (blueprints) for large urban-areas, the evaluation of the 3D reconstructions is not possible. Thus, we use boundary localization as a proxy to reconstruction accuracy and perform the evaluation in 2D. A comparative quantitative analysis of reconstruction accuracies corresponding to different classification accuracies confirms the strong correlation between the two. We present the results which show a consistent and considerable reduction in the reconstruction accuracy. Finally, we present the simple yet compelling concept of latent learning and the implications it carries within the context of deep learning. We posit that a network trained on a primary task (i.e. building classification) is unintentionally learning about auxiliary tasks (e.g. the classification of road, tree, etc.) which are complementary to the primary task. Although embedded in a trained network, this latent knowledge relating to the auxiliary tasks is never externalized or immediately expressed but instead only knowledge relating to the primary task is ever output by the network. We experimentally prove this occurrence of incidental learning on the pre-trained ICT-Net and show how sub-classification of the negative label is possible without further training/fine-tuning. We present the results of our experiments and explain how knowledge about auxiliary and complementary tasks – for which the network was never trained – can be retrieved and utilized for further classification. We extensively tested the proposed technique on the ISPRS benchmark dataset which contains multi-label ground truth, and report an average classification accuracy (F1 score) of 54.29% (SD=17.03) for roads, 10.15% (SD=2.54) for cars, 24.11% (SD=5.25) for trees, 42.74% (SD=6.62) for low vegetation, and 18.30% (SD=16.08) for clutter. The source code and supplemental material is publicly available at http://www.theICTlab.org/lp/2020ICTNet/.

A novel convolutional neural network architecture of multispectral remote sensing images for automatic material classification

For real-world simulation, terrain models must combine various types of information on material and texture in terrain reconstruction for the three-dimensional numerical simulation of terrain. However, the construction of such models using the conventional method often involves high costs in both manpower and time. Therefore, this study used a convolutional neural network (CNN) architecture to classify material in multispectral remote sensing images to simplify the construction of future models. Visible light (i.e., RGB), near infrared (NIR), normalized difference vegetation index (NDVI), and digital surface model (DSM) images were examined.This paper proposes the use of the robust U-Net (RUNet) model, which integrates multiple CNN architectures, for material classification. This model, which is based on an improved U-Net architecture combined with the shortcut connections in the ResNet model, preserves the features of shallow network extraction. The architecture is divided into an encoding layer and a decoding layer. The encoding layer comprises 10 convolutional layers and 4 pooling layers. The decoding layer contains four upsampling layers, eight convolutional layers, and one classification convolutional layer. The material classification process in this study involved the training and testing of the RUNet model. Because of the large size of remote sensing images, the training process randomly cuts subimages of the same size from the training set and then inputs them into the RUNet model for training. To consider the spatial information of the material, the test process cuts multiple test subimages from the test set through mirror padding and overlapping cropping; RUNet then classifies the subimages. Finally, it merges the subimage classification results back into the original test image.The aerial image labeling dataset of the National Institute for Research in Digital Science and Technology (Inria, abbreviated from the French Institut national de recherche en sciences et technologies du numérique) was used as well as its configured dataset (called Inria-2) and a dataset from the International Society for Photogrammetry and Remote Sensing (ISPRS). Material classification was performed with RUNet. Moreover, the effects of the mirror padding and overlapping cropping were analyzed, as were the impacts of subimage size on classification performance. The Inria dataset achieved the optimal results; after the morphological optimization of RUNet, the overall intersection over union (IoU) and classification accuracy reached 70.82% and 95.66%, respectively. Regarding the Inria-2 dataset, the IoU and accuracy were 75.5% and 95.71%, respectively, after classification refinement. Although the overall IoU and accuracy were 0.46% and 0.04% lower than those of the improved fully convolutional network, the training time of the RUNet model was approximately 10.6 h shorter. In the ISPRS dataset experiment, the overall accuracy of the combined multispectral, NDVI, and DSM images reached 89.71%, surpassing that of the RGB images. NIR and DSM provide more information on material features, reducing the likelihood of misclassification caused by similar features (e.g., in color, shape, or texture) in RGB images. Overall, RUNet outperformed the other models in the material classification of remote sensing images. The present findings indicate that it has potential for application in land use monitoring and disaster assessment as well as in model construction for simulation systems.

Efficient and high-performance pedestrian detection implementation for intelligent vehicles

Implementing pedestrian detection real-time embedded systems remains a major challenge. Detecting pedestrians in an advanced driver assistance system requires a lot of time and resources. The method based on Multi-Scale Region Covariance Descriptor (MSRCD) and Support Vector Machine (SVM) is one of the most effective approaches to perform pedestrian detection. However, such implementation is difficult to be executed in real time on embedded systems. This paper presents three improvements to adapt the solution based on the MSRCD descriptor and SVM classifier for embedded pedestrian detection. First, a new feature combination capable to provide the most accurate description at a minimum processing time for MSRCD is proposed. Second, to speed up the SVM classification, a new approach that adopts the mean SVM technique and Euclidian distance is proposed. Third, parallel implementation is exploited to accelerate processing time on multi-core architectures. The software implementation is performed using the INRIA data set. 18.94% processing time speed-up, 48.21% less memory usage, and 2.38% improved detection accuracy are achieved using the proposed descriptor. Moreover, 58.22% processing time speed-up is obtained for the proposed classifier while keeping the same testing accuracy. The parallel implementation is performed using dual ARM Cortex-A9 processors of a selected Xilinx zynq platform. The obtained results confirmed the effectiveness of the proposed parallelization at accelerating the computing time about 3 times the original sequential processing.

Graph Convolutional Networks for the Automated Production of Building Vector Maps From Aerial Images

Building footprint delineation from remote sensing imagery is a basic task in surveying and mapping and geographic information system (GIS), which benefits many engineering applications but requires an enormous amount of human delineation. In this study, we aim to find a way to replace human delineation with automated algorithms. For this purpose, we designed a novel pipeline for the automated production of building vector maps from aerial images, which consists of a bounding-box generation module, a graph convolutional network (GCN)-based polygon prediction module, and an empirical polygon regularization module. First, we introduce the bounding box generation module based on region-based object detection, which along with our overlap cropping strategy is used to generate a bounding box for each building instance. The generated bounding boxes have a horizontal rectangle version and a rotated version, which are used to initialize the next stage of our method. Second, we propose a GCN-based method, which is the core of this study, to conduct the initial accurate building polygon prediction by integrating multiresolution optimization and multilevel loss constraints. Our proposed method, to the authors’ best knowledge, is the first of its kind that introduces GCN to building extraction. The final step is to apply a regularization algorithm to translate the predicted polygons into structured and highly accurate building footprints. We validated the proposed method on two large aerial building data sets, WHU data set and Inria data set, where it was shown to significantly outperform other state-of-the-art methods more than 10%. Specifically, with the ground-truth rotated bounding boxes of buildings, our method is able to automatically delineate 91% of the buildings in the WHU data set and with the predicted rotated bounding the percent reached human-level delineation.

Research on small target pedestrian detection based on improved YOLO

Aiming at the problems of low detection accuracy and speed for small target pedestrian in traffic scenes, the YOLO-SP based on YOLO-v4 is proposed. Firstly, the KITTI and INRIA data sets are used to make the new data set, and k-means algorithm is used to cluster the anchor box. Secondly, in the feature fusion phase (Neck), we increase the number of fused channels and simplify the number of output channels. Finally, to optimise the loss function, GIOU is used to calculate the coordinate loss, Focal is used to calculate the confidence loss. The test shows that all the improvement measures show better effect on small and overlapping pedestrians, the final detection accuracy (AP) is increased by 4.0%, the detection speed is accelerated by 11.3%. YOLO-SP has a certain practicality in the small target pedestrian detection.

Pedestrian detection algorithm based on improved SSD

The algorithm proposed in this paper introduces the deeply separable fusion hierarchical feature model into the backbone network of Single Shot MultiBox Detector (SSD) algorithm, which reduces the complexity of the model by using depthwise Separable Convolution and uses hierarchical structure fusion to enhance features. We add decode-encoder structure between the backbone network and the additional feature layer to propagate the high-level semantic feature downward through decoding, and then integrate the feature with the shallow local detail feature. At the same time, the encoder is used to transfer more advanced semantic features and output the fused shallow features for pedestrian detection, which improves the classification and detection ability of the model. The miss rate of the improved algorithm in this paper is as low as 9.68% on the INRIA data set.