Skip to content

MohamedSebaie/Car_Crashes_Detection_From_Surveillance_Cameras-OrangeLabs-Project

Repository files navigation

Car_Collision_Detection_OrangeLabs

this is car collision detection system me and my team created it as graduation project in ITI ai-pro track undersupervision and mentoring of orange labs

Car Crash Detection can be seen as the Detection of Accident or Not in a video according to the actions occurring in it. It has become one of the most challenging and attractive problems in video classification and detection fields.

The problem itself is difficult to solve by traditional video processing methods because of several challenges such as the background noise, sizes of subjects in different videos, and the speed of Cars.Derived from the progress of deep learning methods, several directions are developed for video detection, such as the violent flow , the long-short-term memory (LSTM)-based model, two-stream convolutional neural network (CNN) model, and the convolutional 3D model. Car Crash Detection is used in some surveillance systems and video processing tools.

Our main problem is Accident Detection which we achieved to solve by using 2 aprroches VIF and 3D Convolution.

Our Pipeline for VIF :

read the camera stream then run our segmentation methodology of each car in the scene then compute VIF for each tracked object and give each feature vector of segmented cars to svm to detect whether this object has collision or not .

alt text

- We use Yolo V4 to detect all vehicles in the scene

- track each detected vehicle using deep sort

- use second step verification using our sift based similarity method to correct tracking errors

- create trajectory of each car and segment it into smaller videos as shown in the image below

alt text

- compute Violent flow for each segmented object

- get the magnitude and angle histogram and combine them into 36 bins

- get the mean of these values over 30 frames and give the results to SVM

the main file to run the project is our colab kernel to run the demo

Our Pipeline for 3D Convolution (3D_ResNet-18):

we used transfer learning on pretrained convolutional 3D models that aim to recognize the motions and actions of Cars and all models use Kinetics-400 dataset for the pretrained part and Vision-based Accident Detection From Surveillance Cameras dataset for the finetuned part.

alt text

Our 3D_Convolution Pipeline

alt text

Pytorch Pretrained Models

All pretrained models can be found in this link. https://pytorch.org/vision/stable/models.html

Instructions to Install our Car Crash Detection Package

Pip Package can be found in this link. https://pypi.org/project/Car-Crash-Detection/

  1. Install:
pip install Car-Crash-Detection
pip install pytube
  1. Download the Finetunned Model Weights
import gdown
url = 'https://drive.google.com/uc?id=1-8TyT7MkAS7LLsRTbO03tDsuuoBM6q1D'
model = 'model_ft.pth'
gdown.download(url, model, quiet=False)
  1. Detect Accident or Not by Pass your Local Video:
from car_crash_detection import CrashUtils
# Run the Below Function by Input your Test Video Path to get the outPut Video with Accident Detection or Not
CrashUtils.crashDetection(inputPath,seq,skip,outputPath,showInfo=False,thresholding=0.75)
  1. Show the Output Video with Detection:
from moviepy.editor import *
VideoFileClip(outputPath, audio=False, target_resolution=(300,None)).ipython_display()
  1. To Start Detect the Accident on Streaming
CrashUtils.start_streaming(streamingURL,thresholding=0.75)

About

Car_Crash_Detection_From_Surveillance_Cameras-PyTorch

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published