Pavel Vyacheslavovich Skribtsov ¹, Pavel Aleksandrovich Kazantsev ¹, Aleksey Vladimirovich Dolgopolov ¹

Affiliations
1 PAWLIN Technologies Ltd, Dubna, Russia

Abstract

Background/Objectives: The main objective of this work is to create a new head-shoulder detection algorithm that is able to perform robustly with arbitrary camera's point of view and in any environment in real time on the CPU. Method: To achieve the objective the developed head-shoulder detection algorithm uses two cascades: HOG-Cascade and Cascade based on automatically extracted features. Deep neural network autoencoder is applied for automatic feature extraction. In this work, the positive examples were extracted from the TUD Multiple View Pedestrian database. The negative images that do not contain people were taken from the Nicta Dataset. We also used positive and negative examples collected from online surveillance cameras. The dataset was divided into training, test and validation. Detection quality was estimated using the Recall and False Positives Per Image (FPPI) metrics. Findings: Our study has shown that the combined use of HOG-features and automatically extracted features allows achieving high quality (Recall: 85%, FPPI: 0.5) and performance (5.5 FPS at CPU) in the problem of head-shoulder detection. The results demonstrate that the proposed method is comparable or slightly outperforms the state-of-the-art methods. Improvements/Applications: In the application of the developed detector to real video, it would be possible to reduce the level of false positives and increase the performance using the motion mask or filter based on the model of the scene.

Keywords

Adaboost, Deep Autoencoders, Head-Shoulder Detection, HOG, Naive-Bayes Classifier

Full Text

   

Aleksey Vladimirovich Dolgopolov ¹, Pavel Aleksandrovich Kazantsev ¹, Nikolay Nikolaevich Bezuhliy ¹

Affiliations
1 Nakhimov Black Sea Higher Naval School, Sevastopol, Russia

Abstract

Background/Objectives: The main objective of this work is the creation of a new algorithm for detecting ships in the high resolution images obtained by unmanned aerial vehicle. Method: To achieve the objective the developed ship detection algorithm uses two cascades: shape-based cascade and cascade based on automatically extracted features. Deep neural network autoencoder is applied for automatic feature extraction. Findings: Images from Google Maps, Yandex Maps, and a small fraction of the images taken from the Internet were used for training and testing the developed approach. In total, the dataset consists of 1000 images containing ships. On average, each image contains 5.1 ships. The dataset was divided into training (50%) and test (50%). Detection quality was estimated using the Recall and Precision metrics. Our study has shown that the combined use of shape-based cascade and cascade based on automatically extracted features allows achieving high quality (Recall: 0.95, Precision: 0.94) and performance in the problem of ship detection by UAV. The results demonstrate that the proposed method is comparable or slightly outperforms the state-of-the-art methods. Improvements/Applications: Applying the detector in real video from the UAV, it would be possible to improve performance by parallelization of this approach on the GPU.

Keywords

Connected Components, Deep Autoencoders, Naive-Bayes Classifier, Ship Detection, UAV.

Full Text

   

Pavel Aleksandrovich Kazantsev ¹, Pavel Vyacheslavovich Skribtsov ¹, Sergey Olegovich Surikov ¹

Affiliations
1 PAWLIN Technologies Ltd., Dubna, RU

Abstract

Objectives: This study is an attempt to develop a fast real-time multi-object tracking algorithm with competitive precision and accuracy in crowded scenes. Methods/Statistical Analysis: This research extends multi-object tracking-by-detection framework by utilizing physical characteristics of trackers and their affinity for tracker guidance, in addition to detections and representation models. Guidance is done by particle filter which weights are updated through reworked observation model, featuring a term based upon tracker affinity. Tracker affinity is also used for tracker grouping that removes redundant trackers following the same target, and triggering a special propagation mode during occlusions that prevent identity switches. Findings: Our research has shown that trackers affinity matrix and algorithm features based on it yield substantial extra accuracy for tracking-by-detection framework, while taking a minor fraction of processing time. Improvements/Applications: In this paper we have introduced an approach that makes it possible to use less accurate, but faster detectors and representation model classifiers, enabling real-time processing, while keeping competitive precision and accuracy.

Keywords

Multi-Object Tracking, Tracking-by-Detection, Particle Filter, Pedestrian Detection, Particle Filtering, Online Learning, Tracker Affinity, Surveillance.

Full Text

   

Pavel Vyacheslavovich Skribtsov ¹, Pavel Aleksandrovich Kazantsev ¹

Affiliations
1 PAWLIN Technologies Ltd., Dubna, Moscow Region, RU

Abstract

Background/Objectives: This paper proposes a new type of regularization for deep learning neural networks that is capable of explicit separation of the lower dimensional hidden layer input pattern representation into two components: class information component and transform component. Methods: Currently, researchers involved in pattern recognition problems are actively searching for the replacement of deterministic feature extraction algorithms by unsupervised methods capable of generating optimal domain-specific image features during the training process of auto-associative multilayer neural networks. The result of the training process of the deep neural network with a “bottleneck” hidden layer is the task-oriented encoder capable of efficient input signal dimensionality reduction. Findings: Many important useful properties of the encoder including the degree of invariance of the feature extraction to input signal transformations (perturbations) greatly depend on the particular form of the regularization applied. In addition to the regular weight decay smoothing component the suggested regularization has two additional components: the first one minimizes the spread of the class-describing features under different pattern transforms and the other component minimizes the spread of the transformation description features for the objects with same perturbations but from the different classes. Class-membership information from the training sequence is used along with the introduced estimator of the similarity of pattern transform to compute the regularization terms. The research reveals that a private case of the suggested regularization corresponds to the well-known Frobenius norm of Jacobian matrix of the encoder activations, therefore the contribution of this paper can be seen as a non-local extension of the encoder Jacobian-based family of deep neural network regularizers embedding invariance to non-local input pattern transformations into the deep neural network feature extraction pipeline. Experiments carried out on the synthetic and real pattern datasets show promising results and encourage further investigation of the proposed approach. Improvements/Applications: This method can be used for areal images recognition invariant to lighting, weather and orientation, for example for the recognition of vehicles and other landmarks in the images obtained by the unmanned aerial vehicles (UAV).

Keywords

Auto Encoder, Deep Neural Networks, Invariant Image Recognition, Regularization.

Full Text