This technology is based on multi-task deep learning model to solve object detection and semantic segmentation problem. The network utilizes the semantic attention module to boost the object detection performance. The whole technology can be used to detect drivable area, lane line and also moving object on the road with single camera. The captured information can be utilized to several advanced driver assistant systems.

1.Multi-task learning model to joint solve object detection and semantic segmentation with novel semantic attention module to boost performance at the little cost of computation.
2.The overall network is pretty light-weight and owns acceptable accuracy. It runs at 10 FPS on NVIDIA Jetson Xavier embedded system and also 15 FPS on Texas Instrument TDA2x.

Multi-task learning network is more efficient because it integrate several algorithm to a single-unified one, and save a lot of data transferring time and memory. And also the computation and number of parameters can dramatically reduce because the network share the same backbone encoder. This technology is suitable for areas that require object detection or semantic segmentation, such as driving purposes, surveillance systems, and more.