Abstract: Fatality on roads is one of the biggest issues due to which people lose their life. The enormous number of participant injuries and fatalities emphasizes the essential necessity for worldwide road safety. Our project focuses on gathering photos, comparing them to the training dataset, and categorizing them as accidents or not accidents. A message is then sent to the nearby hospital along with the coordinates of the accident's location. We employed the Densenet-161 architecture for this project where each layer is connected to all layers next to it. This is done in order to maximize the information flow between network tiers as each layer sends its information gained to all the layers next to it. In contrast to Resnet, it connects features rather than adding features to them to combine features. As a result, the "ith" layer contains I inputs and is made up of all the convolutional blocks' passing features. All other "n-i" layers are split into its own features. Data from on-board cameras is validated using а сомраriсоn with straightforward classifiers that use only video or audio data. This introduces '(n(n+1))/2' connections in the network as opposed to just 'n' connections as in traditional deep learning using a learning system, the learning system is located. The trained algorithm is tested using YouTube clips of related incidents. According to experimental tests, the suggested CAR detection system outperforms various improved classifiers and it offers up to 80% accuracy.
Keywords: CNN, densenet, accidents.
|
DOI:
10.17148/IJARCCE.2022.11791
