Vehicle detection and classification with spatio-temporal information obtained from CNN

Vehicle detection and classification with convolutional neural networks (CNN) is encouraging and fascinating for intelligent transportation applications. In general, CNN is trained with huge labeled vehicle images, which are rare to find in developing country scenarios. This work proposes a vehicle detection technique with CNN that does not require any labeled vehicle dataset. A CNN is trained with road marks and these act as the background. When a vehicle occupies a road mark a logic 1 is recorded else 0 is saved in the database. This accumulated occupancy data is spatio-temporal information from which vehicle width and length are approximated and classified. This method does not require any GPU for training as well as for implementation in real-time. The detection accuracy of the proposed method on real-time videos is about 98.5%. The runtime performance of CNN model on NVIDIA Jetson Tx2 is 0.01 s, whereas on RaspberryPi4 is 0.03 s. The given method even works in shadow and low-light conditions.