Multiple Objects Detection and Classification for Street Intersection Surveillance Video Based on Deep Learning

EasyChair Preprint 790

Abstract

This research uses deep learning techniques to classify image objects of intelligent image recognition from various public environments (e.g., intersections, campuses and community safe). After receiving an image, able to use quickly pre-trained several categories such as large cars, small cars, motorcycles, and bicycles to classification, using the “you only look once” deep learning architecture for training and detection. We filter and balance classes and quantities of input training data to ensure they can better model the images and improve detection stability. Therefore, the mAP of our balanced dataset category quantity detection result from data input through object selection improved from the original 80.36 to 90.26 . The advantages of this technology are real-time detection and statistical benefits. In addition to reduced labor costs, our intelligent detection reduces the probability of accidents.

Keyphrases: Intersections, YOLO, large vehicle, self created dataset, training data, vehicle classification, vehicle detection

@booklet{EasyChair:790,
  author    = {Ri-Chen Feng and Daw-Tung Lin and Yi-Yao Lin},
  title     = {Multiple Objects Detection and Classification for Street Intersection Surveillance Video Based on Deep Learning},
  howpublished = {EasyChair Preprint 790},
  year      = {EasyChair, 2019}}