Specialized Car Detector for Autonomous Driving

Fast and accurate perception of the environment are important for autonomous vehicle research, especially the detection of other vehicles in the front. Different from object detection tasks in other fields where high accuracy and real-time detection speed are required, the model size of object detection system applied to autonomous vehicles should be as small as possible due to the limitations of vehicle-mounted embedded system. The object detection system of the autonomous vehicle takes too many pedestrians and other factors into account, which lead to a degradation in detection performance. In this work, we propose the so-called SpCarDet, a specialized car detector, which is a one-stage object detection network. Based on the Fire module, Stacked feature map and Shortcut connections, a new and small model size backbone network is built. We also present the bounding boxes generation network named AGnet that is optimized for car detection. The real-time accurate detection of the vehicle is achieved, and the model size of network is small to enable embedded system deployment. Experiments on the KITTI dataset demonstrate the effectiveness of our algorithm.