Visual tracking of dynamic defective contour based on fused long short-term memory model

For the rapid detection of high-speed objects in a compact dynamic environment, visual simultaneous localization and mapping (VSLAM) may fail to extract defective contours caused by occlusion. Combined with the long short-term memory (LSTM) model, a fast and accurate method is proposed to detect and track dynamic defective contour. Specifically, to improve the perception ability of the LSTM model, visual geometry group (16) (VGG16) and Canny edge detection are balanced to extract high-dimensional spatial features and low-dimensional edge features, respectively. These features are linearly concatenated to form multimodal inputs. Next, by suppressing defective objects caused by the jitter of bounding box, an overlap bounding box perceptual method combined with Kalman filtering is introduced to realize the cross-frame tracking efficiency of dynamic object from low to high. Finally, to address devices occlusion in the feeding system under varying working conditions, a contour transfer approach in the temporal occlusion domain is introduced, leveraging Lucas-Kanade optical flow to enhance contour continuity. Experiments are conducted on simulated videos of motion coordination between stamping stations and handling manipulators in a periodic high-speed stamping production line. Results demonstrate that the proposed method achieves stable and efficient visual tracking of defect contour in dynamic environments.