Integrated detection and tracking for ADAS using deep neural network

The recent advancements in computer vision technology have ensured that it has an increasingly important position in intelligent transportation. This paper proposes an integral system, including object detection and tracking, to recognize multiple objects in dynamic and complex real-world scenes. A backbone network of the single shot multi-box detector (SSD) is implemented using an improved SqueezeNet for performance improvement. The object detector is followed by an online object tracker that fuses multiple information features, including the appearance feature extracted by CNNs, motion information, and shape information. Both the detector and tracker can well balance accuracy and processing time. The proposed system shows acceptable performance, especially the detector demonstrates the best performance among real-time models on the KITTI test benchmark.