Fast Segmentation-Based Object Tracking Model for Autonomous Vehicles

On-road object tracking is a critical module for both Advanced Driving Assistant System (ADAS) and autonomous vehicles. Commonly, this function can be achieved through single vehicle sensors, such as a camera or LiDAR. Consider the low cost and wide application of optical cameras, a simple image segmentation-based on-road object tracking model is proposed. Different from the detection-based tracking with bounding box, our model improves tracking performance from the following three aspects: 1) the Positional Normalization (PONO) feature is used to enhance the target outline with common convolutional layers. 2) The inter-frame correlation of each target used for tracking relies on mask, this helps the model reducing the influences caused by the background around the targets. 3) By using a bidirectional LSTM module capable of capturing timing correlation information, the forward and reverse matching of the targets in consecutive frames is performed. We also evaluate the presented model on the KITTI MOTS (Multi-Object and Segmentation) task which collected from out door environment for autonomous vehicle. Results show that our model is three times faster than Track RCNN with slightly drop on sMOTSA, and is more suitable for deployment on vehicular low-power edge computing equipment.