3.1 STM32 Embedded Motion Target Tracking and Detection Method

3.1.1 STM32 architecture

An embedded system is application centric and built on a computer foundation. The customizable software and hardware configuration is equipped with dedicated microprocessors, focusing on the design of customized software, decoders, and simulators for traditional computer use ^[13]. Its hardware includes high-performance microprocessors and various interface circuits, while its software covers real-time operating systems and application software. STM32 is an ARM Cortex-3 core microcontroller launched by STMicroelectronics, which has the characteristics of high performance, low cost, and low power consumption ^[14]. The study adopts a hardware control platform based on STM32 microprocessor, combined with motion target detection and tracking algorithm for interactive software design. Real time control of the camera is achieved through the rotating platform of the mechanical transmission system to complete MTTD. Fig. 1(a) shows the functional modules of the system. Fig. 1(b) is a schematic diagram of the mechanical rotation part studied.

Fig. 1. Functional module diagram of the system and schematic diagram of mechanical rotation part.

Fig. 2. JTAG interface schematic diagram.

In Fig. 1(a), the main working modules of the control platform hardware system include power supply, LCD screen, OV7725 camera, JTAG interface, and motor drive control. A 3.2-inch LCD screen with a resolution of 320x240 is selected as the display screen, and data are stored using the Flexible Static Memory Controller (FSMC) on the STM32 ^[15]. The LCD controller undergoes a series of conversions to convert the collected signal data into RGB format and write it into the Graphics Random Access Memory (GRAM). In Fig. 1(b), the entire control module is lightweight and compact, and the mechanical rotation system operates stably. Due to the weight and installation requirements of the camera, only one flange type deep groove ball bearing is used for fixation in both horizontal and vertical directions. Fig. 2 is theJTAG interface.

In the JTAG interface schematic in Fig. 2, control and communication are achieved through four control lines (Test Mode Select: TMS, Test Clock: TCK, Test Data In: TDI, Test Data Out: TDO). TCK receives the test clock input, TDI is used for data input, TDO outputs data from the JTAG port, and TMS sets the JTAG port to a specific test mode.

3.1.2 Introduction to existing PCNN

Unlike traditional pulsed NN and ANN, PCNN simulates information transmission between neurons through pulses ^[16,17]. Fig. 3 shows the structure of PCNN neurons.

Fig. 3. Structure of PCNN neurons.

In Fig. 3, the pulse coupled neuron model consists of three main parts: receiving domain, modulation, and pulse generation. The reception threshold consists of a linking channel and a feedback channel. The receiving domain receives input and feedback information from adjacent neurons. The feedback channel not only receives information between adjacent neurons, but also receives external stimuli, represented by Eq. (1).

(1)

$ F_{ij} [n]=F_{i} {}_{j} [n-1]e^{-aF} +V_{F} \sum M_{ijkl} Y_{kl} [n-1] +I_{ij} . $

In Eq. (1), the ignition information is represented by $Y_{kl} $. The time decay constant is $a_{F} $. The external input signal is represented by $I_{ij} $. $M$ is the connection weight matrix. The feedback of the $(i,j)$-th neuron is $F_{ij} [n]$. Eq. (2) is the operational relationship of synaptic input information, used to describe the reception of input between adjacent neurons.

(2)

$ L_{ij} [n]=L_{i} {}_{j} [n-1]e^{-aL} +V_{L} \sum M_{ijkl} Y_{kl} [n-1] . $

In Eq. (2), $V_{L} $ represents the feedback constant. The linear input of the connected neuron of the $(i,j)$-th neuron is $L_{ij} [n]$. The internal activities are represented by Eq. (3).

(3)

$ U_{ij} [n]=(1+\beta L_{ij} [n]F_{ij} [n] . $

In Eq. (3), $\beta $ represents the modulation constant between neurons. Whether neurons emit pulses depends on the relationship between internal activity $U_{ij} [n]$ and threshold $E_{ij} [n]$, represented by Eq. (4).

(4)

$ Y_{ij} [n] =step(U_{ij} [n]-E_{ij} [n-1])\\ =\left\{\begin{aligned}1, U_{ij} [n]>E_{ij} [n-1],\\0,\text{others}.\end{aligned}\right. $

In Eq. (4), the ignition function $Y_{ij} $ represents neuronal ignition when it is equal to 1. When it is equal to 0, it indicates no ignition. The mathematical expression of $E_{ij} [n]$ is represented by Eq. (5).

(5)

$ E_{ij} [n]=E_{ij} [n-1]e^{-aE} +V_{E} Y_{ij} [n-1] . $

In Eq. (5), $V_{E} $ represents the threshold constant. $aE$ is the time decay constant. $E_{ij} $ represents the threshold for change.

3.1.3 OFGA-based PCNN parameter optimization

To address the slow modeling and poor detection performance of traditional algorithms, an optimized PCNN background differential motion target detection algorithm is proposed. It adopts the dual threshold idea to simplify PCNN and improve it, while introducing the Optimal Family Genetic Algorithm (OFGA). Fig. 4 is a simplified flowchart of the PCNN parameter optimization algorithm based on OFGA.

Fig. 4. Flowchart of PCNN parameter optimization algorithm based on OFGA.

In Fig. 4, the PCNN parameter optimization algorithm based on OFGA first receives image input, and then initializes it through the OFGA system, using the PCNN system to calculate fitness.Traditional PCNN relies on experience in initial parameter settings, and parameter selection may be poor, which can affect convergence speed and performance. OFGA-PCNN utilizes genetic algorithms to automatically optimize initial parameters, avoiding the uncertainty of manually set parameters and quickly finding the global optimal parameter combination, thereby accelerating convergence. Genetic algorithms have strong global search capabilities and can find optimal solutions in complex search spaces. Through genetic algorithms, OFGA-PCNN can avoid falling into local optima, thereby improving the global optimal performance of the network. In addition, genetic algorithms can efficiently explore the search space and accelerate the training process of the network through selection, crossover, and mutation operations. In this experiment, it is determined that the termination condition is met. If the condition is met, the image is output. Otherwise, genetic operation is performed and a new population is generated. During this process, the key parameters of PCNN are decoded to achieve optimization. After simplification, the mathematical expression of $F_{ij} [n]$ is represented by Eq. (6).

(6)

$ F_{ij} [n]=I_{ij} . $

In Eq. (6), $I_{ij} $ represents the external input stimulus signal. The linear input $L_{ij} [n]$ of the connected neurons of the $(i,j)$-th neuron is expressed mathematically using Eq. (7).

(7)

$ L_{ij} [n]=\sum V_{L} W_{ijkl} Y_{kl} [n-1] . $

$Y$ represents whether the neuron ignites in Eq. (7). $V_{L} $ is a connection constant. The mathematical expression of $Y_{ij} [n]$ is represented by Eq. (8).

(8)

$ Y_{ij} [n]=\left\{\begin{array}{l} {1{\rm \; \; U}_{ij} [n]>E_{ij} [n-1]} \\ {0{\rm \; \; others}} \end{array}\right. $

In Eq. (8), ${\rm U}_{ij} [n]$ represents the feedback input of the $(i,j)$-th neuron. $E_{ij} [n-1]$ is a threshold function, and its threshold is crucial. Usually, traditional algorithms use a single fixed threshold and set $E_{ij} [n-1]$ as a constant. However, a threshold that is too small can lead to significant noise, while a threshold that is too large may result in incomplete motion detection results. Therefore, the dual threshold idea is introduced into each iteration of PCNN, and an improved PCNN motion detection algorithm is proposed. The dual threshold idea uses upper and lower thresholds in binary threshold segmentation ^[18].

3.2 Motion Target Tracking and Detection Method on the Foundation of Pulse-Coupled Neural Network

The research on target tracking algorithms is committed to achieving real-time tracking of detected moving targets. By confirming the shape, contour, or position of the moving object in each frame of the image, target tracking calculates the direction and trajectory of the object's motion, and provides timely feedback on the object's future motion status. Mean Shift is a common target tracking algorithm ^[19]. Fig. 5 shows the Mean Shift vector.

Fig. 5. The mean shift vector.

In Fig. 5, the red dot is $x$. The dark blue center points around are sample point $x_{i} $. The arrow represents the offsetting vector of $x_{i} $ relative to $x$. The average offset points towards the direction of the densest sample points, which is the gradient direction. Mean Shift includes two parts: target representation and target localization. The user selects the target area and candidate window, and establishes a feature histogram to represent the target model and candidate model. In motion target tracking, the target model is determined, the similarity with the candidate model is calculated, and the transfer vector is obtained. By iteratively calculating the Mean Shift vector, the target converges to its true position, completing the update of the target position and achieving target tracking ^[20]. In the feature space, the center of the target region is the origin, and the center of the candidate region in subsequent frames is another position. Therefore, the feature vectors of the target model are represented by Eq. (9).

In Eq. (9), $m$ represents the dimension of the feature vector. To conduct research, it should establish a large number of motion target models for training. Eq. (10) shows the library of input samples.

In Eq. (10), $g_{c,n} $ represents the $n$-th action segment. The clip contains $m$'s movements and postures. The calculation of $g_{c,n} $ action film in the $m$-th posture is represented by Eq. (11).

After a complete action segment is projected to the output space, a ``trajectory'' will be formed in the output space and a set of index numbers containing timing information will be obtained. Eq. (12) is the specific calculation.

In Eq. (12), $O_{c,n} $ is used to identify the index sequence of targets for each category. According to the histogram statistics rule, a histogram of an action sequence containing $n$ poses is represented by Eq. (13).

In Eq. (13), $f$ represents the frequency of the occurrence of the $u$-th output node in the action. $N$ is the quantity of postures included in the action. New input actions are computed by matching Euclidean distances to known action templates to discriminate the class of unknown actions. The difference in action is calculated by the normalized inner product of the feature vectors of two poses, represented by Eq. (14).

In Eq. (14), $f_{i,k} $ and $f_{j,k} $ are the $k$-th eigenvector values of attitudes $p_{i} $ and $p_{j} $, respectively. $f_{k} (\max )$ is the $k$-th feature's maximum value. $f_{k} (\min )$ represents this feature's minimum value. $w_{k} $meansits weight. The common frame is the foundation of the algorithm, used to ensure that the front and back actions have similar poses in order to achieve smooth transitions. According to our understanding, for any two pose data, Eq. (15) can be used to measure the distance between their center of gravity.

In Eq. (15), $(x,y,z)$means the gravity center's coordinates of the human body. Conduct real-time tracking experiments based on the traditional Mean Shift tracking algorithm and the optimal family optimized PCNN motion target detection algorithm. Fig. 6 shows the entire algorithm process.

Fig. 6. Flow chart of detection algorithm.

In Fig. 6, first, Mean Shift is used to represent and locate the initial target, and feature histograms of the target model and candidate model are established. Subsequently, the targeting and candidate models' similarity is calculated using the optimal family optimized PCNN, and the transition vector is obtained. By iteratively updating the Mean Shift vector, the target quickly and accurately converges to its true position, thereby achieving real-time tracking and detection of moving targets.

4.1 Performance Verification of Target Tracking and Detection Methods

When conducting performance experiments on video sequences, the experimental computing environment was an Intel Core i5-2450 2.5GHz processor, 4G DDR3 memory, running on a 64 bit Microsoft Windows 7 operating system, and a 1G NVIDIA GT630M graphics card. The software environments used included VS2010, Intel C++, and OpenCV2.4.8. The video was collected from the school laboratory and contains a total of 300 frames.

Fig. 7. Dataset example.

As shown in Fig. 7, From the perspective of a single image, the video exhibits the following characteristics: the laboratory serves as the background, and the human target has a high degree of separability from the background. Some color features of the characters are similar to the background, such as the dark stripes on the clothes and the computer in the back, and the white stripes on the walls. The person is close to the camera and there is no obstruction, and the target's movement speed is moderate. To further explore the performance gain of OFGA optimization algorithm in PCNN motion detection, the additional computational complexity brought by this algorithm was considered in standard surveillance video experiments. The study adopted the GPU universal computing architecture, further introduced the NVIDIA GTX580 2G graphics card, and equipped it with VS2008 and CUDA4.0 software environments. Fig. 8 shows the evolution curves of GA and OFGA algorithms for solving PCNN parameter optimization problems.

Fig. 8. Evolution curves of GA and OFGA algorithms for solving PCNN parameter optimization problems.

In Fig. 8, the two algorithms used the same parameter settings (pm $=0.75$, P $=0.00$). According to Fig. 8(a), the number of dominant families in OFGA was 10, and the fitness value was the average value of the algorithm after 20 iterations. According to Fig. 8(b), the ordinary GA had the problem of slow convergence speed and easy falling into premature (i.e., local optimal solution). The improved OFGA had greatly improved this issue. By searching in the micro space around dominant individuals, OFGA achieved significant acceleration in the early stages of evolution and successfully subdivided the search space, greatly reducing the probability of the algorithm falling into premature convergence. Precision and success rate were used as two evaluation metrics to assess the performance comprehensively, which was compared with four excellent algorithms on the OTB platform. The detailed information of the algorithm used is shown in Table 1.

All experiments were conducted as a single evaluation, with an overlap threshold of 0.5 and a center position error threshold of 20. Fig. 9 shows the comparison of accuracy and success rate between the proposed algorithm and other algorithms.

Fig. 9. Comparisonresults of different algorithms.

In Fig. 9, DSSTPF exhibited excellent performance. In Fig. 9(a), its accuracy reached 86.6%. In Fig. 9(b), its success rate was 81%. Compared with the other four algorithms that performed better in OTB2015, DSSTPF showed a significant improvement in overall data. Compared to DSST, thisresearchmodel's accuracy and success rate had been improved by 12.7% and 14%, respectively. So it has better performance compared to other algorithms. The ROC curve of the model was analyzed, and the results are shown in Fig. 10.

Fig. 10. Comparison of ROC curves of models.

Fig. 10(a) shows the ROC curve of the SOTA model, and Fig. 10(b) shows the ROC curve of the model proposed in this study. As shown in Fig. 10, the area under the curve of the proposed model is significantly larger than that of the SOTA model. The research results indicate that the proposed model has excellent performance.Table 2 shows its accuracy and success rate compared to the other four algorithms (fDSST, DSST, KCF, Struck) on different attributes.Three datasets were used, namely the Channel and Color Cluttered dataset, the Low Resolution dataset, and the Blurred and Cluttered dataset.

According to the data in Table 2, the proposed improved algorithm performs well in all aspects. In these cases, its accuracy and success rate are significantly higher than other algorithms. Overall, the algorithm proposed by the research institute has achieved satisfactory tracking results, although there is room for improvement under certain specific attributes.

4.2 Performance Verification of Target Tracking and Detection Methods under Different Influences

To further investigate the adaptability to lens motion or background disturbances, Table 3 lists six commonly used tracking models and the F1 Mean values of the proposed model. This value effectively reflects the algorithm's ability to extract salient targets and suppress non-salient targets.

In Table 3, this model performed excellently in target tracking when facing camera motion or background disturbances. It is worth noting that the proposed method demonstrated better performance compared to various other methods in three different video scenes. For the Walking database, this method had also shown superior performance in various methods besides GBVS. Fig. 10 shows a comparison of algorithm accuracy and success rate under gear attributes.

Fig. 11. Comparison of algorithm accuracy and success rate under block attribute.

According to Fig. 11(a), the proposed algorithm maintained an accuracy of 89.1% in the presence of occlusion. According to Fig. 11(b), the proposed algorithm maintained success rates of 89.1% and 83.4% in the presence of occlusion, which had the best tracking performance compared to the other four algorithms. The study used the Human3.6M dataset for training, which contained approximately 3.6 million annotated human motion data samples and their corresponding RGB images. The entire dataset can be divided into 11 major categories, each consisting of 15 subcategories. These major categories mainly represent 11 different professional model experimenters, while the subcategories mainly cover different human movements. Fig. 12 shows the results obtained by three different models using average error as the evaluation metric in the Human3.6M dataset.

Fig. 12. Comparison of identification errors of different methods on Human3.6M dataset.

According to Fig. 12(a), except for A, the proposed method had the smallest error compared to other research methods proposed by Du's et al., with values of 118, 95, 96, 137, and 112 on B, C, D, E, and F, respectively. According to Fig. 12(b), except for d, the proposed method had the smallest error compared to other methods proposed by Du's and others, with errors of 99, 120, 115, 117, and 112 on a, b, c, e, and f, respectively.Compare the real-time performance and computational resource consumption of various models. The results are shown in Table 4.

In Table 4, A represents real-time performance and B represents computational resource consumption. As shown in the table, with the increase of time, the model proposed in this study has consistently achieved high real-time performance and low computational resource consumption. Excelling in all performances. The experimental results show that the proposed model performs well among various models.