Develop the finance app of the company
Develop business automation software
Develop the REST api web application and depolyment on the cloud
Develop the AI chat bot which answer the question about the securities.
Development enviroment:
Badminton player motion identification algorithm with deep learning and computer vision Sep 2020 - Sep 2022
Responsibilities: Data collection, writing code, testing results
Task: Enter a badminton match video and identify the movements of the players appearing in the video.
Optimization: Created a more general data set, and reduced bias through data enhancement. Through pre-knowledge, it
was proposed to use opposing athletes and based on the baseline.
Features such as the speed and direction vector of badminton use OpenPose, TrackNet and transfer learning methods to
extract feature vectors and filter the background and
Other useless information greatly reduces the dimension of the feature vector and simplifies the amount of calculation.
Results: The accuracy of action detection increased from 46.7% in the baseline to 86.7%, which greatly improved the
accuracy.
Technology stack: pytorch, python
PSO algorithm to solve TSP problem Oct 2020 - Feb 2021
Responsibilities: Code writing and testing
Background: TSP is a combinatorial optimization problem. Because of the large amount of calculation, it tends to be solved
using heuristic algorithms, such as the ant colony algorithm. This time I hope to try the PSO algorithm
Solve the TSP problem, but the simple PSO algorithm is used to solve continuous space problems and cannot solve the
discrete domain TSP problem.
Task: Improve the PSO algorithm so that it can solve the TSP problem.
Optimization: Combine the generated answer with the crossover and mutation operations of the genetic algorithm, and
exchange part of the solution with a certain probability. If this solution is worse than the original solution, then the solution is
solved again.
In line with the probability acceptance criterion of the simulated annealing algorithm, the inferior solution is accepted with a
certain probability to jump out of the local optimal solution. Introduced checking for path intersections and unintersections
For the cross algorithm, it has been proven that the crossed path must be inferior to the non-crossed solution, which greatly
speeds up the algorithm fitting speed.
Results: It solved the problem that the traditional PSO algorithm, which can only handle continuous space problems, cannot
be used for discrete TSP, and optimized the iteration speed of the algorithm.
Technology stack: C++
Competition Experience&HONORS & AWARDS
Third Prize in the ACM-ICPC 2018.4
MCM/ICM American College Student Mathematical Modeling Competition SP Award 2018.2
Rank 8th in the world in IEEEXtreme 11.0 Extreme Programming Competition 2017.10
Rank 39th in the world in IEEEXtreme 12.0 Extreme Programming Competition 2018.10
Third Prize in NOIP programming competition 2013.11
Second Prize in NOIP programming competition 2014.11
Third Prize in NOIP programming competition 2015.11
Second Prize in NOIP programming competition 2016.11
Third Prize in the FIRST LEGO League Robot League Competition 2013.4
Second Prize in the FIRST LEGO League Robot League Competition 2014.4
qwucJ6RQ202311231243
qwucJ6RQ202311231243
qwucJ6RQ202311231243
qwucJ6RQ202311231243