Presentation Title

Object detection for body worn video

Faculty Mentor

Bao Wang

Start Date

18-11-2017 10:00 AM

End Date

18-11-2017 11:00 AM

Location

BSC-Ursa Minor 87

Session

Poster 1

Type of Presentation

Poster

Subject Area

engineering_computer_science

Abstract

The author applies Faster-RCNN, a neural-network based object detection algorithm, to the field of Body-worn video with the goal of differentiating between police and non-police individuals. Along with creating a functional object detector, this paper also attempts to improve Faster-RCNN results. To achieve this goal, The author has two steps. First, the author applies image repair techniques and LDMM to de-noise the training image sets, the dataset that the author have manually prepared and labelled. Second, the author tries multiple approaches to enhance the network performance. In recognition of inter-class distance and intra-class variance, the author introduces Center Loss and Contrastive Loss functions to the existing Faster-RCNN loss function. The author also introduces gradient-boosting (GB) to the original detection pipeline to build a stronger classifier. The author has demonstrated that noise in the natural images actually increases detection accuracy and gradient-boosting technique has improved the detection result. The continuing research in improving object detection algorithm for police detection is in progress.The topic is interesting and has a potential to generate high impact.

This document is currently not available here.

Share

COinS
 
Nov 18th, 10:00 AM Nov 18th, 11:00 AM

Object detection for body worn video

BSC-Ursa Minor 87

The author applies Faster-RCNN, a neural-network based object detection algorithm, to the field of Body-worn video with the goal of differentiating between police and non-police individuals. Along with creating a functional object detector, this paper also attempts to improve Faster-RCNN results. To achieve this goal, The author has two steps. First, the author applies image repair techniques and LDMM to de-noise the training image sets, the dataset that the author have manually prepared and labelled. Second, the author tries multiple approaches to enhance the network performance. In recognition of inter-class distance and intra-class variance, the author introduces Center Loss and Contrastive Loss functions to the existing Faster-RCNN loss function. The author also introduces gradient-boosting (GB) to the original detection pipeline to build a stronger classifier. The author has demonstrated that noise in the natural images actually increases detection accuracy and gradient-boosting technique has improved the detection result. The continuing research in improving object detection algorithm for police detection is in progress.The topic is interesting and has a potential to generate high impact.