Smart surveillance system for identifying bikers without helmets using deep learning
MetadataShow full item record
Modern world is progressing quickly along with technology and one of the major sectors is transportation technology. Day by day the number of people are increasing and the number of vehicles are increasing too. As a result we have easy access to vehicles these days but at the same time it has increased the number of road accidents. Among the various types of road accidents, motorcycle accident is one of the main type which causes severe injuries and in some cases death. The only protection motorcyclists use is their helmet. Most countries has a law on wearing helmet otherwise it will be punished, but many people often break this law and as a result it increases the percentage of severe injury and death. In a populated country it's hard to keep track of the bikers who don't use helmets because of the huge number of bikes moving at a time. In this circumstances, we have developed a solution which can identify the bikers who don't use helmets. Our system uses image processing and deep Convolutional Neural Networks(CNNs) which is used to identify who breaks the law of wearing helmet, with helmet vs without helmet identification and finally motorcycle license plate recognition. We have used multiple models of object identification and evaluated in terms of speed vs accuracy. The results we have got indicates that due to the increase of law enforcement and awareness of the traffic police, the use of helmet has decreased a bit than before but there are still a lot of bikers who don't use helmets on regular basis. As we are going to detect people without helmet and then will keep the registration number from the license plate, we can easily identify the law breakers and the government can punish them accordingly. We have used the Tensor ow library for our system. The models we have used in our system are SSD Mobilenet v2 and Faster RCNN inception v2. For SSD Mobilenet V2 the accuracy for the helmet was 90 percent, human was 55 percent, bike was 80 percent and number plate was 95 percent. This model is quite light and we can use this trained model even in mobile devices. For Faster RCNN inception v2 it took more computations but the accuracy we got was slightly better as it is more heavyweight than SSD Mobilenet. The accuracy we got for this was 92 percent for the helmet, 58 percent for the human, 81 percent for bikes and 96 percent for number plates.