Intelligent Driving Division - Artificial Intelligence Dept.Task: Pre-research of point cloud segmentationResult: Build a neural network to segment point cloud. Achieved 73% mIoU on test set.

This is an Independent Research Project of ME Dept. of UMich, advised by Prof. Huei Peng, director of Mcity. The object of the project is to produce a 2D/3D lane graph from a single image by end-to-end deep learning architecture. Specifically, it introduces Graph Neural Network combined with VGG16 on lane detection, trained with graph-structured road dataset based on tuSimple dataset.The advantage of this approach is that by applying graph based deep learning, the network can fully utilize the graph nature of lanes with a more explainable physical-enhanced algorithm design, as well as generating an omnipresent representation for lanes in variant scenarios as highways, intersections, etc.

This is a competition launched by Prof. Ram Vasudevan from ME Dept.Competition Object:To generate vehicle control signals and predicted trajectory on a virtual track few miles long. The vehicle model is based on Pacejka ‘Magic’ nonlinear model with 6 degrees of freedom. After submitting the algorithm, the system would randomly generate thirty obstacles along the track, and detect whether the vehicle would run into obstacles based on the control inputs. Afterwards, the system would compare the time used by the algorithm to complete all outputs, and by the vehicle to finish the track.My Approach:My algorithm generated the control signals by combining MPC and Sampling based motion primitive methods. First, the algorithm created the reference trajectory offline first, with only the track boundaries (obstacles unknown). Then, it produced the obstacle-avoidance trajectory by considering online computation for obstacles within 700 feet based on the reference trajectory. Sampling based method was applied on straight or low curvature part of the track, while MPC was applied on curvy part where nonlinearity is prominent.Hightlights:1. Designed a regional convex polygon selection algorithm for quadratic programming, to ensure the global optimal solution in selected areas.2. Combining MPC and Sampling idea to ensure the computation can be concentrated on sharp turning and where obstacles are dense, where MPC implemented.3. The algorithm gives control outputs within 0.6 miles at a time. Meanwhile, the actual running of the vehicle has little deviation from the calculated one in the end point under the nonlinear model.

Summer Research Internship in PATH (DeepDrive Project) Institute, University of California, Berkeley in summer 2017. My job was assisting Postdoctoral Mr. Chuang to collect and train dataset for a self-driving model car, and tested it on PATH field. The self-driving model car utilize SqeezeNet, and successfully performed autonomous driving.