Monocular Vision-Based Real-Time Vehicle Detection at Container Terminals

We present a practical approach to vehicle detection at container terminals based on a single camera and prevailing convolutional neural network models in computer vision domain. Aiming at container terminal scenarios, we introduce a specialized data labelling strategy for network training, as well as an optimized setting of crucial hyperparameters, leading to a significant improvement on results. Our solution achieves 83% precision with 90% recall for semitrailer trunks within 30 m ahead of the vehicle-mounted monocular camera, at a speed of 32 frames per second (FPS) on a Nvidia Titan X for 416 x 416 image input, also providing more alternatives of fairly easy speed/accuracy trade-off. Compared to traditional lidar-based vehicle detection method for autonomous driving, our solution is more robust for particular container terminal scenarios while still maintaining a real-time performance by completely eliminating the region proposal stage.