Multiple object tracking and trajectory recognition based on deep learning

In automobile plastic parts factory, the workers need to check every automotive plastic assembly whether the fastening buckle is missing. This manual method is inefficient. To solve this problem, the space and channel attention mechanism is applied to YOLOv3-tiny network, which uses K-means clustering to find the best suitable size of anchor box. This algorithm has significantly improved detection accuracy in fastening buckles in comparison with original YOLOv3 algorithm, and the speed has increased by 3 times. Threshold segmentation and Hu comparison method are adopted to correct the confidence of the detection target with weight, so as to filter out the false target. Finally, the Lucas-Kanade optical flow method and the Delay-precision mechanism are used to realize the tracking and recognition of the fastening buckle of the automobile plastic assembly. In the experiment of fastening buckle video sequence, the MOTA(Multiple-Object Tracking Accuracy) of this algorithm is about 3.9%~7.6% higher than other algorithms. In 200 experiments, the recognition accuracy of automotive plastic components and fastening buckles reaches 100%.