Pedestrian detection (pedestrian Detection) resource consolidation

first, the paper
Articles in the review category:
[1] D. Geronimo, and A. M.lopez. vision-based pedestrian Protection Systems for Intelligent vehicles, book, 2014.
[2] P.dollar, C. Wojek,b Schiele, et al pedestrian detection:an evaluation of the state of the art [J]. IEEE transactions on Patternanalysis andmachine Intelligence, 2012, 34 (4): 743-761.
[3] Cogs Zhi, Li Shaozi, Chen Shuyuan, etc. Survey of pedestrian detection technology [J]. Journal of Electronics, 2012, 40 (4): 814-820.
[4] M. Enzweiler, and D.gavrila. Monocular pedestrian Detection:survey and experiments [J]. IEEE Transactions on Pattern analysis Andmachine Intelligence, 2009, 31 (12): 2179-2195.
[5] D. Geronimo, A. M.lopez and A. Sappa, et al Survey of pedestrian detection for advanced driverassistance systems [ J]. IEEE transactionson Pattern Analysis and Machine Intelligence, 2010, 32 (7): 1239-1258.
[6] Jia Huising, Zhang Yujin. A survey of pedestrian detection based on computer vision in vehicle-assisted driving system [J], Journal of Automation, 2007, 33 (1): 84-90.
[7] Xu Yanwu, Cao Xianbin, Qiao Hong. New progress and key technology prospect of pedestrian detection system [J], Electronics Journal, 2008, 36 (5): 368-376.
[8] Duyutan; Chen Feng; Xu Wenli; Li Yongbin; A review of human motion recognition based on vision, electronic newspaper, 2007. 35 (1): 84-90.

[9] Zhu Wenjia. Research on the key technology of pedestrian detection based on machine learning [D]. The first chapter, Master's degree thesis, Shanghai Jiaotong University. 2008. Instructor: Shi.

2014_eccv_30hz Object Detection with DPM V5
2014_eccv_knowing a good HOG filter whenyou see it:efficient selection of filters for detection
2014_eccv_unsupervised dense objectdiscovery, Detection, Tracking and reconstruction
2014_eccv_comparing salient objectdetection Results without Ground Truth
2014_ECCV_RGBD salient Object Detection:abenchmark and algorithms
2014_eccv_saliency Detection with Flash andno-flash Image Pairs
2014_eccv_object Co-detection via Efficientinference in a fully-connected CRF
2014_eccv_context as supervisory signal:discovering Objects with predictable Contex
2014_eccv_object Detection and viewpointestimation with auto-masking neural Network
2014_eccv_deep learning of Scene-specificclassifier for pedestrian Detection
2014_eccv_spatio-temporal Object Detectionproposals
2014_eccv_strengthening the effectivenessof pedestrian Detection with spatially pooled Features
2014_eccv_video Object Discovery andco-segmentation with extremely Weak supervision
2014_eccv_face Detection without Bells andwhistles
2014_eccv_human Detection using Learnedpart Alphabet and Pose Dictionary
2014_eccv_spade:scalar Product Acceleratorby Integer decomposition for Object Detection
2014_eccv_towards Unifiedobject Detection and segmentation
2014_eccv_depth Based Object Detection frompartial Pose estimation of symmetric Objects
2014_eccv_training deformable Object modelsfor Human Detection based on Alignment and clustering
2014_eccv_spatio-temporal Matching Forhuman Detection in Video
2014_eccv_facial Landmark Detection by Deepmulti-task Learning
2014_eccv_latent-class Hough forests for 3DObject Detection and Pose estimation
2014_eccv_gis-assisted Object Detection andgeospatial localization
2014_eccv_sliding Shapes for 3D Object detectionin Depth Images
2014_eccv_integrating Context and Occlusionfor Car Detection by hierarchical And-or Model
2014_eccv_simultaneous Detection andsegmentation
2014_eccv_learning Rich Features from Rgb-dimages for Object Detection and segmentation
2014_its_toward real-time pedestrian detection based on a deformable template model
2014_pami_scene-specific Pedestrian detection for static video surveillance
2014_cvpr_pedestrian Detection in low-resolution Imagery by learning multi-scale intrinsic Motion structures (Mims)
2014_cvpr_switchable deep Network for pedestrian Detection
2014_cvpr_informed haar-like Features Improve pedestrian Detection
2014_cvpr_word Channel Based multiscale pedestrian Detection without Image resizing and Using only one Classifier
2013_bmvc_surveillance camera autocalibration based on pedestrian height distribution
2013_virtual and real world adaptation for pedestrian detection
2013_search space reduction in pedestrian detection for driver assistance system based on projective geometry
2013_cvpr_robust multi-resolution pedestrian Detection in traffic Scenes
2013_cvpr_optimized pedestrian Detection for multiple and occluded people
2013_cvpr_pedestrian Detection with unsupervised and multi-stage Feature learning
2013_cvpr_single-pedestrian Detection aided by Multi-pedestrian Detection
2013_cvpr_modeling Mutual Visibility Relationship in pedestrian Detection
2013_cvpr_local Fisher discriminant analysis for pedestrian re-identification

Second, DataSets

MIT Database
The database is an earlier public pedestrian database with a total of 924 pedestrian images (ppm format, width height of 64x128), and shoulder-to-foot distances of about 80 pixels. The database contains only front and back two views, no negative samples, no training set and test set. Dalal, such as the use of "HOG+SVM", the database on the detection accuracy rate of nearly 100%.

Inria Database
The database is currently the most used static pedestrian detection database, providing the original image and the corresponding callout file. The training set has a positive sample of 614 (including 2,416 pedestrians), a negative sample of 1218, the test set has a positive sample 288 (including 1126 pedestrians), negative sample 453. Most of the human body in the picture is standing posture and height is greater than 100 pixels, some of the labels may be incorrect. The image is mainly from GRAZ-01, personal photos and Google, so the picture is high definition. Some training or test images under XP operating system cannot be seen clearly, but can be read and displayed normally by OPENCV.

Daimler Pedestrian Database
The database is captured by a car camera and is divided into two data sets for detection and classification. The training sample set of the test data set has a positive sample size of 18x36 and 48x96 images of each 15560 (3915x4), the minimum height of the pedestrian is 72 pixels, and a negative sample of 6744 sheets (size 640x480 or 360x288). The test set is a video of about 27 minutes (640x480 resolution), a total of 21790 images, including 56,492 pedestrians. The classification database has three training sets and two test sets, each with 4800 pedestrian images, 5000 non-pedestrian images, all sizes of 18x36, and 3 additional non-pedestrian photo sets, each with 1200 images.

Caltech Pedestrian Database
    The database is now a large pedestrian database, taken by car camera, about 10 hours, video resolution of 640x480,30 frames per second. It is labeled about 250,000 frames (about 137 minutes), 350,000 rectangles, 2,300 pedestrians, and also labels the time correspondence between rectangles and their occlusion. The dataset is divided into Set00~set10, where set00~set05 is a training set and set06~set10 is a test set (callout information is not yet exposed). There are three methods of performance evaluation: (1) Training with external data, testing in Set06~set10, (2) 6-fold cross-validation, selecting 5 of them for training, another doing test, adjusting parameters, and finally giving the performance on the training set; (3) Using set00~set05 training, Set06~set10 do the test. Because the label information for the test set is not exposed, it needs to be submitted to Pitor Dollar. The result submission method is a test for every 30 frames, and the results are saved in the TXT document (the file is named I00029.txt I00059.txt ...). Each TXT file indicates that a pedestrian is detected in the format "[Left, top,width, height, score]". If no pedestrians are detected, the TXT document is empty. The database also provides the corresponding MATLAB toolkit, including the video annotation information reading, the ROC (Receiver operatingcharacteristic Curve) curve diagram and non-maximum value suppression tools.

Tud Pedestrian Database
In order to evaluate the function of motion information in pedestrian detection, TUD Pedestrian database provides image pair to calculate optical flow information. The positive sample of the training set is 1092 pairs of images (image size is 720x576, including 1776 pedestrians), negative samples are 192 pairs of non-pedestrian images (handheld camera 85 pair, car camera 107 pair) and 26 images taken by car camera (including 183 pedestrians) as additional training sets. The test set has 508 pairs of images (the time interval of the image pair is 1 seconds, with a resolution of 640x480) and a total of 1326 pedestrians. Andriluka has also built a database to verify their detection and tracking combined with the pedestrian detection technology. The training set of the dataset provides information about the rectangles of the pedestrian, the size and location of the split mask and its various parts (feet, calves, thighs, torso, and head). The test set is 250 images (including 311 fully visible pedestrians) to test the performance of the detector, and 2 video sequences (Tud-campus and tud-crossing) are used to evaluate the performance of the tracker.

NICTA Pedestrian Database
The database is the current large-scale static image pedestrian database, 25551 pictures with single, 5207 high-resolution non-pedestrian pictures, the database has been divided into training sets and test sets, convenient comparison of different classifiers. Overett such as "Realboost+haar" evaluation of the training samples of the translation, rotation and aspect ratio of the impact of various factors on the classification performance: (1) Pedestrian height of at least 40 pixels, (2) at low resolution, for Haar characteristics, The performance of increasing sample width is better than that of increasing sample height, (3) the size of training picture is greater than the actual size of pedestrian, that is, background information is helpful to improve performance, and (4) to translate the training sample to improve the detection performance, rotation has little effect on the performance improvement. The above conclusions are of great significance to the construction of pedestrian database.

ETH Pedestrian Database
ESS constructs a pedestrian database based on binocular vision for multi-person pedestrian detection and tracking. The database uses a pair of on-board AVT Marlins f033c Camera for shooting, the resolution is 640x480, the frame rate 13-14fps, give the calibration information and pedestrian labeling information, the depth of information using the confidence of the propagation method obtained.

CVC Pedestrian Database
The database currently contains three datasets (CVC-01, CVC-02 and Cvc-virtual), which are mainly used for pedestrian detection studies in vehicle-assisted driving. CVC-01[GERONIMO,2007] There are 1000 pedestrian samples and 6,175 non-pedestrian samples (from non-pedestrian images in the road area of the picture, unlike the non-pedestrian samples of the pedestrian database as natural images of the sky, sandy beaches and trees). The CVC-02 contains three sub-datasets (CVC-02-CG, cvc-02-classification, and Cvc-02-system) for three different tasks for pedestrian detection: generation, classification, and system performance assessment for the area of interest. Image acquisition using BUMBLEBEE2 stereo color vision system, resolution 640x480, focal length of 6mm, the distance from the camera 0~50m pedestrians, the smallest pedestrian picture is 12x24. CVC-02-CG mainly for the production of candidate regions, there are 100 color images, including depth and 3D point information; cvc-02-classification mainly for pedestrian classification, training set has 1016 positive samples, 7650 negative samples, The test set is divided into the classification based on the cutting window (570 pedestrians, 7500 non-pedestrians) and the whole picture (250 images containing pedestrians, 587 pedestrians); Cvc-02-system is mainly used for system performance evaluation, including 15 video sequences (4364 frames) and 7,983 pedestrians. Cvc-virtual is a virtual pedestrian dataset generated by the Half-Life 2 image engine, which contains 1678 virtual pedestrians and 2048 non-pedestrian images for testing.

USC Pedestrian Database
The database contains three sets of Datasets (Usc-a, Usc-b, and usc-c) that provide callout information in XML format. USC-A[WU, 2005] The picture from the network, a total of 205 pictures, 313 standing pedestrians, pedestrians do not exist between each other, shooting angle of the front or back; Usc-b's images are mainly from the Caviar video library, including pedestrians from various viewpoints, some of which are obscured by pedestrians, A total of 54 pictures, 271 pedestrians, usc-c 100 pictures from the network, 232 pedestrians (multi-angle), pedestrians do not have a mutual occlusion.

Third, Source Code
1.INRIA Object Detection and Localization Toolkit, Dalal in 2005 proposed a pedestrian detection method based on hog features, one of the classic articles in the field of pedestrian detection. Hog features are also being used in other areas such as target detection and recognition, image retrieval and tracking.
2. Real-time pedestrian Detection. Jianxin Wu implements a fast pedestrian detection method.
3. Hough transfom for pedestrian Detection. Olga Barinova, CVPR paper:on detection of multiple object instances using Hough Transforms
4. HIKSVM, HOG+LBP+HIKSVM, the classic method of pedestrian detection.
5. Groundhog, gpu-based Object Detection with geometric Constraints, In:icvs, 2011. Cuda version of HOG+SVM, video.
6.100fps_pds, pedestrian detection at frames per second, R. Benenson. CVPR, 2012. Real-time (⊙o⊙) Oh. Real-time!!!
7. Pom:probabilistic occupancy Map. Multiple camera pedestrian detection.
8. Pitor Dollar Detector.  Integral Channel Feature + multi-scale feature approximation + multi-feature fusion. real-time!