Single-filter CNN for Vehicle Recognition

According to the World Health Organization, over 1.35 million people are killed on roads every year. Initiatives like the European Vision Zero aim to eliminate road fatalities and injuries. To increase road safety, vehicle manufacturers equip vehicles with multiple sensors and provide drivers with advanced driver-assistance systems (ADAS). Processing a continuous stream of data from multiple sources in real-time can be computationally demanding and increase power consumption. Offloading these complex tasks to edge or cloud computing can enhance the speed of inference and save power, which is particularly crucial for electric cars. However, to avoid network congestion caused by the transmission of large amounts of data, sensor data should be filtered or compressed within the vehicle itself. We have proposed a lightweight method to filter frames containing vehicles, ensuring a transparent decision-making process. This method based on region proposals and binary classification using a single convolutional neural network (CNN) kernel, achieves an overall accuracy of 0.81.