US10719940B2 - Target Tracking Method And Device Oriented To Airborne-Based Monitoring Scenarios - Google Patents

Target detecting and tracking are two of the core tasks in the sphere of visual surveillance. Relu activated totally-linked layers to derive an output of 4-dimensional bounding field knowledge by regression, wherein the four-dimensional bounding box data includes: horizontal coordinates of an higher left nook of the primary rectangular bounding box, vertical coordinates of the higher left nook of the first rectangular bounding field, a size of the primary rectangular bounding box, and iTagPro technology a width of the primary rectangular bounding box. FIG. Three is a structural diagram illustrating a goal tracking device oriented to airborne-based mostly monitoring situations in line with an exemplary embodiment of the current disclosure. FIG. Four is a structural diagram illustrating another goal tracking device oriented to airborne-primarily based monitoring situations based on an exemplary embodiment of the present disclosure. FIG. 1 is a flowchart diagram illustrating a goal tracking methodology oriented to airborne-primarily based monitoring eventualities in keeping with an exemplary embodiment of the present disclosure. Step 101 acquiring a video to-be-tracked of the goal object in actual time, and performing body decoding to the video to-be-tracked to extract a first frame and a second body.



Step 102 trimming and capturing the first body to derive an image for first interest region, and trimming and capturing the second frame to derive a picture for target template and an image for second curiosity region. N times that of a size and width data of the second rectangular bounding box, respectively. N could also be 2, that's, the length and width data of the third rectangular bounding field are 2 times that of the size and width information of the primary rectangular bounding box, respectively. 2 times that of the original information, acquiring a bounding field with an area 4 times that of the unique knowledge. Based on the smoothness assumption of motions, it is believed that the place of the goal object in the primary body must be found in the curiosity area that the realm has been expanded. Step 103 inputting the image for goal template and the picture for first interest area into a preset appearance tracker network to derive an look tracking position.



Relu, and the number of channels for outputting the characteristic map is 6, 12, 24, 36, 48, and 64 in sequence. Three for the remaining. To make sure the integrity of the spatial place data in the characteristic map, the convolutional community doesn't embrace any down-sampling pooling layer. Feature maps derived from completely different convolutional layers in the parallel two streams of the twin networks are cascaded and integrated using the hierarchical function pyramid of the convolutional neural community while the convolution deepens continuously, respectively. This kernel is used for performing a cross-correlation calculation for dense sampling with sliding window kind on the characteristic map, which is derived by cascading and integrating one stream corresponding to the image for first interest area, and a response map for appearance similarity can be derived. It can be seen that in the appearance tracker community, the monitoring is in essence about deriving the position the place the target is situated by a multi-scale dense sliding window search within the interest area.



The search is calculated based on the goal look similarity, that's, the appearance similarity between the target template and the picture of the searched position is calculated at every sliding window position. The place the place the similarity response is massive is very probably the position where the goal is located. Step 104 inputting the picture for first curiosity region and iTagPro device the image for second curiosity area into a preset movement tracker network to derive a movement monitoring position. Spotlight filter body difference module, a foreground enhancing and background suppressing module in sequence, iTagPro technology whereby every module is constructed based on a convolutional neural network structure. Relu activated convolutional layers. Each of the number of outputted function maps channel is three, wherein the function map is the distinction map for the input image derived from the calculations. Spotlight filter frame distinction module to obtain a frame difference movement response map corresponding to the interest regions of two frames comprising previous body and subsequent body.



This multi-scale convolution design which is derived by cascading and secondary integrating three convolutional layers with completely different kernel sizes, aims to filter the movement noises caused by the lens motions. Step 105 inputting the appearance tracking position and the motion monitoring place right into a deep integration network to derive an built-in closing monitoring position. 1 convolution kernel to revive the output channel to a single channel, thereby teachably integrating the tracking outcomes to derive the ultimate monitoring place response map. Relu activated absolutely-related layers, and a four-dimensional bounding field knowledge is derived by regression for outputting. This embodiment combines two streams tracker networks in parallel in the process of monitoring the target object, wherein the target object's appearance and movement data are used to perform the positioning and iTagPro technology monitoring for the target object, and the final tracking position is derived by integrating two times positioning information. FIG. 2 is a flowchart diagram illustrating a target monitoring technique oriented to airborne-primarily based monitoring eventualities according to a different exemplary embodiment of the present disclosure.