Download PDFOpen PDF in browser

Customized Deep Learning Technique for Vehicle Detection Along with Speed Estimation

EasyChair Preprint 7851

21 pagesDate: April 28, 2022

Abstract

The method proposed over here in this paper is a vehicle speed estimation technique on moving vehicle under the cctv camera surveillance. For real-time vehicle detection, the YOLO (You Only Look Once) technique is employed, and the centroid approach is used to estimate vehicle speed. The video frame is converted to grayscale so that it may be processed by the computer as 0 and 1. The brightness of the scale is represented by each number. Then, by looking at these statistics, we train the YOLO Convolutional Neural Network to learn to identify the final detection. YOLO reframes object recognition as a single regression issue by taking the entire image and going directly from image pixels to bounding box coordinates and class probabilities. The next step is to compute bounding boxes (boxes that encompass the objects) using IoU (Intersect over Union) and NMS (non-maximum suppression). The IoU indicates how closely the machine's predicted bounding box fits the bounding box of the real item. However, because of the process, a problem of over-identification with a specific object arises. NMS ensures that the best cell is found among all these bounding boxes. Rather than concluding that a single car in the image has numerous causes, NMS chooses the boxes with the highest likelihood of determining the same vehicle. The vehicle centroid values are calculated after the cars have been detected. The distance traveled by vehicle is calculated using the centroid value. The speed of the vehicle is calculated after sorting out the distance that has been covered by the vehicle. YOLO is an effective and efficient strategy that epitomizes the spirit of machine learning in the suggested methodology for our vehicle recognition and speed estimation system. YOLO initially trains with 416*416 photographs, then retrains for 30 epochs at a 10-3 learning rate using 416*416 images. After training, the classifier has a top-one accuracy of 99.4% and a top-five accuracy of 99.3%.

Keyphrases: CNN, IOU, NMS, Regression, YOLOv5, bounding box, centroid, deep, speed estimation, vehicle detection

BibTeX entry
BibTeX does not have the right entry for preprints. This is a hack for producing the correct reference:
@booklet{EasyChair:7851,
  author    = {Benjamin Tamang and Sanskar Poudel and Sagar Bhandari and Bipin Damase and Sagar Pande},
  title     = {Customized Deep Learning Technique for Vehicle Detection Along with Speed Estimation},
  howpublished = {EasyChair Preprint 7851},
  year      = {EasyChair, 2022}}
Download PDFOpen PDF in browser