Too Far to See? Not Really! — Pedestrian Detection with Scale-aware Localization Policy

Proposed System: an active pedestrian detector that explicitly operates over multi-layer representation and localization policy.

Multi-layer representations of ResNet are respectively utilized to compile pedestrian proposals of different sizes, which are then passed to our localization policy module to produce the final outputs.

Existing System: Meanwhile, initial pedestrian proposals are attained by faster R-CNNs techniques, i.e. region proposal network and follow-up region of interesting pooling layer employed right after the specific ResNet convolutional layer of interest, to produce joint predictions on the bounding-box proposals’ locations and categories (i.e. pedestrian or not). 

Methodology:


 A high-level overview of our approach is displayed in Figure, which consists of two main stages. An input image will first pass through an initialization stage to make ready the multi-layer feature representations and the initial pedestrian proposals (of ResNet are respectively utilized to compile pedestrian proposals of different sizes), they are then fed into an active detection stage where a dedicated localization policy (final outputs) is engaged to produce the final bbox predictions by executing sequences of coordinate transformation actions.

ResNet: An input image is passed through the ResNet layers to form multi-layer feature representations, where bbox proposals of different scales are also generated. The region of interest (RoI) pooling layer of is then utilized to pool the feature maps of each pedestrian proposal into a fixed-length feature vector, which is fed into a fully connected layer. At last, each pedestrian proposal ends up with two outputs, one accounts for classification score, and the other one predicts the bbox position 

Localization Policy: Localization Policy (delivering the final results) that is learned from data by exploiting both local contextual information and multi-layer representations using deep reinforcement learning and recurrent neural network (RNN) techniques. Localization Policy that identifies the bounding-box of an object of interest, as well as by for digit recognition with a recurrent attention mechanism: where at each time step t, both spatial and temporal contextual information are incorporated to decide current transformation action.

Drawback: further address the challenged of detecting heavily occluded pedestrian instances, as well as investigating the ability of our approach in tackling more generic object detection scenarios.

Requirement: ResNet, RNN inbuild/plugin supporting IDE MATLAB/ Java/DotNet.

For additional details comment below with requirements.

Comments

Post a Comment

Popular posts from this blog

Secure Data Group Sharing and Dissemination with Attribute and Time Conditions in Public Cloud

Image
 Paper Synopsis: Data dissemination: Dissemination takes on the theory of the traditional view of communication, which involves a sender and receiver. The traditional communication view point is broken down into a sender sending information, and receiver collecting the information processing it and sending information back, like a telephone line. Example: Alice (owner, share)      è       Bob (view, share(with policies)) ( disseminator , re-encrypting the ciphertext) Existing system: Attribute-based encryption Attribute-based encryption (ABE) is one of new cryptographic mechanisms used in cloud to reach flexible and fine-grained secure data group sharing [4]. Especially, ciphertext-policy ABE (CP-ABE) allows data owners to encrypt data with an access policy such that only users whose attributes satisfy the access policy can decrypt the data [5]. The proxy re-encryption (PRE) scheme [7] in a manner could achieve efficient data dissemination in cloud by re-encr
Read more

Lightweight Fine-Grained Search over Encrypted Data in Fog Computing

Image
Aim: propose a Lightweight Fine-Grained Search (LFGS) system which supports fine Grained access control and single keyword search. Existing System: ·          the existing ciphertext retrieval schemes rarely focus on the fog computing environment and most of them still impose high computational and storage overhead on resource-limited end users. ·         existing CP-ABKS schemes are approximately proportional to the complexity of access policy, which greatly hinder the uses of resource-limited IoT devices. ·          existing CP-ABE schemes is that encryption or decryption operations involve time-consuming pairing operations and a number of other operations increasing with the complexity of access policy . Proposed System: Lightweight Fine-Grained ciphertexts Search (LFGS) system in fog computing by extending Ciphertext-Policy Attribute-Based Encryption (CP-ABE) and Searchable Encryption (SE) technologies, which can achieve fine-grained access control and keyword sea
Read more

A Comprehensive Study on Social Network Mental Disorders Detection via Online Social Media Mining

Image
Aim: To automatically identify SNMD patients at the early stage according to their OSN data with a novel tensor model that efficiently integrate heterogeneous data from different OSNs. Existing System: The SNMD data from different OSNs may be incomplete due to the heterogeneity (drawback) . For example, the profiles of users may be empty due to the privacy issue, different functions on different OSNs (e.g., game, check-in, event), etc.  Proposed System : propose a machine learning framework, namely, Social Network Mental Disorder Detection (SNMDD) , that exploits features extracted from social network data to accurately identify potential cases of SNMDs. The paper also exploits multi-source learning in SNMDD and propose a new SNMD-based Tensor Model (STM) to improve the accuracy.   Advantage: propose a novel tensor-based approach to address the issues of using heterogeneous data and incorporate domain knowledge in SNMD detection. Objectives: ·          to explore data m
Read more