An Improved Multi-objective Visual Tracking Algorithm

An improved real-time multi-target tracking algorithm based on the combination of improved YOLOV3-Tiny and deep-sort is proposed. The target tracking environment and target are simple, so it is easier to use the self-built data set to fit and test the vehicle it belongs to. In order to prevent the phenomenon of over-fitting, the affine transformation algorithm is used to expand and transform the data set of heavy trucks. After adding the self-built data set, it has a better effect on the original data set test. By replacing the main network of Yolov3-Tiny with a more concise neural network structure and using the effective combination of Kalman filter and Hungarian algorithm, the whole structure is more portable and fast, easier to deploy to small computers even mobile terminals. Compared with Faster RCNN, MobilenetV2, YOLOV3-Tiny and other algorithms on self-built data sets and Pcal VOC data sets, the proposed algorithm has higher real-time performance and accuracy, and robustness. After the neural network is improved, the accuracy of the proposed algorithm is improved by more than 2% compared with the original algorithm. And the computing speed is faster, and the network is more portable.