Search

The symptoms of the same tomato disease in different stages are obviously different, and different diseases show some similarities. Traditional pattern recognition methods can not reflect the dynamic changes of the pathological characterization, and the practicability is poor. To solve this problem, this paper proposed a Multi-Scale AlexNet recognition model for mobile platform based on convolutional neural network (CNN), and implemented a tomato leaf disease image recognition system for agricultural workers based on Android. Many parameters and large memory utilization of traditional AlexNet model are unfit for mobile platform, this paper adjusted the network structure of the traditional model by removing the local response normalization(LRN) layer, modifying the full connection layer, setting up different convolution kernel extraction features, designed a multi- scale recognition model based on the AlexNet. The model consists of 6 layers. It can optimize the training time and memory utilization and achieve high precision. After removing the LRN layer, there was a 30% decrease in running time. Extending the single convolution kernel into multi-scale (1 × 1,3 × 3,5 × 5,7 × 7) convolution kernel then fused at the first layer, removing full connection layer 6 and 7, and taking the place of global average pooling layer, then the model size was only 30.2 M. The forward propagation rate (FRP) and backward propagation rate (BPR) were reduced, and the global average pooling is better than the global maximum pooling on recognition accuracy. So the Multi-Scale AlexNet model used global average pooling in the 5th layer. In image preprocessing phase, in order to avoiding over fitting of the trained model caused by the unbalanced distribution of sample numbers, we had zoomed, flipped, color jittering, add noise and rotated the original pictures of dataset randomly to get the augmented dataset, and used 70% of the pictures as the train dataset and the rest as the validation dataset(20%) and test dataset(10%). These pictures were quantized to 224 × 224 dpi for Multi-Scale AlexNet training, and the original dataset and augmented dataset were used to train models. In order to validate the performance of the proposed model, comparative tests were done between Multi-Scale AlexNet and traditional pattern recognition method. It repeated 600 tests. The results showed that the CNN model achieved 92.7%, the high average recognition accuracy of each disease and each disease in the early, middle and late stages. Compared with the other CNN Net model(MobileNet, SequeezeNet, LeNet-5), the Multi-Scale AlexNet achieved the highest recognition accuracy, and reached 95.8% on the late stage disease dataset, but the SequeezeNet model used less memory. The MobileNet and SequeezeNet model reached lower recognition accuracy on the middle and late stages dataset, that because their convolution size was small. The recognition system was implemented on Android platform, and then test was done on field dataset. The results showed that the average recognition accuracy was 89.2%, its less value was due to the complex background of image. Then the system can meet the requirements of disease image recognition on mobile platform in production practice. The research results provide a method for disease image recognition based on convolution neural network, and provide a reference for automatic identification of crop diseases and engineering applications. [ABSTRACT FROM AUTHOR]

The maternal behaviors reflect the health and welfare of the sows, which directly affect the economic benefit of the pig farm. Computer vision provides an effective, low-cost and non-contact method for monitoring the behavior of animal for precision farming. Under the scene of piggery, it is a challenge for 24-hour automatic recognition of lactating sow postures due to the daily illumination variations, influence of heat lamp, and adhesion between piglets and sows. This paper proposed an automatic recognition algorithm of lactating sow postures based on improved Faster R-CNN (convolutional neural network) using depth video images. To improve the recognition accuracy and satisfy the real-time need, we designed a ZF-D2R (ZF with deeper layers and 2 residual learning frameworks) network by introducing residual learning frameworks into ZF network. First, 3 convolutional layers were added in the ZF network to design ZF-D (ZF with deeper layers). Then, in ZF-D network, shortcut connections were used to form 2 residual learning frameworks. The whole network made up the ZF-D2R network. Moreover, the Center Loss was introduced to Fast R-CNN detector to construct a joint classification loss function. With the joint supervision signals of F-SoftmaxLoss and Center Loss in Fast R-CNN detector, a robust model was trained to obtain the deep feature representations with the 2 key learning objectives, which led to intra-class compactness and inter-class dispersion as much as possible. So, the joint supervision of F-SoftmaxLoss and Center Loss could reduce recognition errors caused by the similar features between different postures. By taking ZF-D2R as basic net and adding the Center Loss to Fast R-CNN detector, the improved Faster R-CNN was built. Experiments to obtain the actual data set of lactating sow posture from the depth video of sows in the 28 pens were performed. The data set included 2 451 standing images, 2 461 sitting images, 2 488 sternal recumbency images, 2 519 ventral recumbency images and 2658 lateral recumbency images. And 5 000 images were randomly chosen as the testing set. The rest of the images were used as training set. To enhance the diversity of training data, dataset augmentation including rotating and mirroring was employed. Based on the Caffe deep learning framework, our improved Faster R-CNN was trained with end-to-end approximate joint methods. By adding 2 residual learning frameworks to ZF-D, the ZF-D2R model improved the MAP (mean of average precision) by 1.28 percentage points. After introducing the Center Loss supervision signal, the MAP of the optimal model reached 93.25%, obtaining an increase of 1.3 percentage points, and the MAP of the method proposed achieved 93.25%. And APs (average precisions) of the 5 classes of postures i.e. standing, sitting, sternal recumbency, ventral recumbency and lateral recumbency were 96.73%, 94.62%, 86.28%, 89.57% and 99.04%, respectively. The MAP of our approach was 3.86 and 1.24 percentage points higher than that of Faster R-CNN based on ZF basic net and Faster RCNN based on the deeper VGG16 basic net, respectively. Our method processed images at a speed of 0.058 s per frame, 0.034 s faster than Faster R-CNN based on VGG16. Our proposed method could improve the recognition accuracy and simultaneously ensure the real-time performance. Compared with DPM (deformable part model) detector plus CNN posture classifier, the MAP of the end-to-end recognition method proposed in this paper was increased by 37.87 percentage points, and the speed was raised by 0.855 s per frame. Our method can be used for the 24-hour recognition of sow behaviors and lays the foundation for the analysis of sow dynamic behavior by video. [ABSTRACT FROM AUTHOR]

Body shape parts acquisition and recognition from moving cattle are difficult to realize automatically especially with the similar color and texture frames or images. Usually, the cattle’s abnormal behavior is hardly detected by human because abnormal behavior detection needs continuous observation, but we can solve this problem with our method. In this paper, we use the Microsoft Company's smart vision sensor (named Kinect) to collect the information of 2 modals from cattle movement (the depth information and RGB information, DRGB). The depth information can be acquired by the infrared sensor and the depth value means the distance between the object and the sensor. The RGB information can be acquired by the normal camera. Based on the color modal information, we propose a randomized nearest neighbor pixel comparison (RNNPC) method to snatch at continuous action frames without static frames. By comparing the distinction of pixels between 2 adjacent frames, we can judge whether there is obvious or micro actions appearing in these 2 neighboring images. And if the distinction value is more than the threshold, the camera would record the cattle’s continuous movements or actions automatically. Based on the depth modal information, we can calculate the mean value of depth which is obtained in the dynamic region by the RNNPC method. And with the mean value of depth, we can filter and transform the invalid pixels into dark pixels in the continuous frames. Meanwhile, the intact shape of the cattle in the original picture is preserved by our method (save the pixels with no change). From the results of filtration, we can see that the original image background area is reduced, and the method (we use SelectiveSearch in this paper) also directly reduces the number of candidate regions compared with the conventional recognition algorithm. According to the training samples scale, the network performance and the characteristic of the key parts of cattle from continuous frames, we adjust the parameters of the deep convolutional neural network, and we set the iterating upper limit and finish the network training when the number of iterating reaches this limit. Finally, with the trained network, we can realize the recognition of key parts from moving cattle in the processed continuous frames. The experimental results show that RNNPC method can save 72% of storage space, and the ratio of valid data for the other 38% of continuous frames can reach 94%. By filtering the invalid pixels, the invalid background information or objects of the original images can be almost removed, and the number of candidate objects generated by the conventional object recognition algorithm can be reduced by an order of magnitude. Compared to the original image, 90% candidate regions are reduced with SelectiveSearch algorithm on DRGB image. By adjusting network parameters, we can improve the convergence speed of deep convolutional neural network in training the samples with similar color and texture, and the single training iteration time does not significantly increase. The average classification accuracy of the net can reach 75.88%, and the image processing rate is 4.32 FPS, and under the same condition, the effect is better than that by the original Fast RCNN. By using the method mentioned above, we can realize body shape parts acquisition and recognition from moving cattle. [ABSTRACT FROM AUTHOR]

Traditional methods for classifying crop pests are based on the appearance of the pests such as their color, morphology and texture using algorithms s back propagation neural network and support vector machine. These methods are sensitive to the environments where the pests appear. Also, the imbalance between training sample numbers, difference in scales at which the training samples are taken, along with that the pests might seasonally change their color and shape, make it difficult for these algorithms to detect and recognize the pests. In order to improve the accuracy of detecting crop pests appearing in diverse environments, this paper proposed a deep convolutional neural network method by combining the spatial pyramid pooling with the improved YOLOv3 deep convolutional neural network algorithm. The proposed method first located the pests on a test image and then identified the species they belongs to. Using the improved network structure in YOLOv3, the proposed method used up-sample and convolution operations to calculate the de-convolution. Bilinear interpolation was used to enhance the output of the network, and the depth characteristics of the pests were extracted by the depth residual neural network. As a result, the improved network can effectively detect and recognize small pests in the images. By fusing with the spatial pyramid pooling, the Yolov3-SPP network can map the extracted eigenvectors to different spatial dimensions. The network can be used at various scales to detect pests of different sizes. Results from identifying and testing 20 types of pests collected by traps showed that the average accuracy of the proposed method was 88.07%, with a detection speed of 26 frames/s. Compared with the typical YOLOv3 algorithm, the proposed methods improved accuracy by 2.8 percentage points. We also reconstructed the maps of the network at each stage and compared them with the typical YOLOv3 algorithm in attempts to demonstrate that the features extracted by the proposed method was more recognizable. The results revealed that the typical YOLOv3 algorithm could miss the targeted objects, while the proposed method is not only able to detect pests of different sizes, it also has high recognition accuracy. Using the same dataset, we compared the proposed method with other detection algorithms such as HOG+SVM, Faster R-CNN, YOLOv3. The comparison showed that the proposed method was 19.61 percentage points higher than that of HOG + SVM, and 9.64 percentage points higher than that of Faster R-CNN, Faster R-CNN was unable to extract small pests, and its recognition accuracy was only 78.43%. [ABSTRACT FROM AUTHOR]

Recognition of larval shrimp is the key to realize the quantification of the behavior of larval shrimp. Recently, the studies on aquatic animal behavior recognition mainly use the typical features method or frame differential method. It is hard to accurately identify larval shrimp due to the lack of key profile information caused by the strong motion distortion and background noise. To solve this problem, we took the Penaeus vannamei as research object, and proposed a recognition method for moving larval shrimp based on the improved PCA (principal component analysis) and Adaboost algorithm in this paper. Firstly, we extracted and optimized the effective features. The improved PCA algorithm was applied to the larval shrimp images with the size of 100 × 100 pixels to reduce the image dimension and extract the external features. We then analyzed and determined the number of principal components according to the distribution of the contribution rate. After that we expounded the theory of the Adaboost algorithm, and used these principal components to build different weak classifiers. A total of 100 images of larval shrimp and background samples were used for training. During the training every weak classifier was selected automatically, and a stage classifier was generated after 20 iterations. To verify the effectiveness of the proposed algorithm, the recognition experiment was performed with 150 sample images containing 100 images of larval shrimp and 50 sample images of background. The images were captured in the Key Laboratory of Fishery Equipment and Engineering Technology of Ministry of Agriculture in Shanghai, China. The experiment results showed that, the recognition rate for stationary and slowly-moving larval shrimp and fast-moving larval shrimp was 100% and 96% respectively, and the overall recognition rate reached 98%. Among all the test samples, the recognition time for each image ranged from 25 to 35 ms, and the average time was 27.898 ms. Compared with the conventional typical feature method and the frame differential method, the proposed method can be used to recognize the moving larval shrimp. Furthermore, the proposed method has better performance, because it effectively restrains the influence of complex situation such as the velocity and direction of motion and the sample's size. This paper provides a technical basis for the larval shrimp's behavior quantification. [ABSTRACT FROM AUTHOR]

Automatic classification of insect species in field crops such as corn, soybean, wheat, and canola is more difficult than the generic object classification because of complex background in filed and high appearance similarity among insect species. In this paper, we propose an insect recognition system on the basis of advanced sparse coding and spatial pyramid model. We firstly learn features from a large amount of unlabeled insect image patches to construct an over-complete dictionary. The sparse coding of insect image patches is obtained by encoding over the dictionary. To enhance discriminative ability of the sparse coding, we then apply multiple scales of filters coupled with different spaces. Finally, the multiple space features of sparse coding are seamlessly embedded into a multi-kernel framework for robust classification. Traditionally, insect recognition has mainly relied on manual identification by expert entomologists. However, for laymen without a thorough understanding of the terminology of insect taxonomy and morphological characteristics, it is hard to discriminate insect categories at the species level. Therefore, effective identification of insects is a key issue that needs to be well addressed. To improve the recognition accuracy, we develop an insect recognition system using advanced sparse coding, spatial pyramid model and multiple-kernel learning techniques. Different from traditional feature representation, a novel feature representation that is multiple-space sparse coding of insect objects is proposed by this work. The work flow of our method can be decomposed into 2 stages. The first stage focuses on image or insect object representation. At this stage, the features of insect images are extracted using advanced sparse coding and spatial pyramid model. The second stage, which deals with effective fusion of multiple insect-categorization features, constructs a kernel-level fusion classifier using all the sparse coding features. At the first stage, for an insect image given, the features of insect images are extracted firstly. The features are then represented as a linear combination of the corresponding training feature dictionary. Then, a multiple-space sparse coding with spatial pyramid model is used to represent insect image in a joint sparse way over all the features. In this process, for the object image, an over-complete dictionary with unlabeled insect images is learned first. Then, the local image patches of the insect object are represented by their sparse codes with the training dictionary. Despite the fact that appearance is modeled using local patches, the global structure information is necessary for accurate insect identification. Consequently, insect appearance is represented by concatenating the location and orientation sparse-coding of all image patches. To obtain the more compact representations of insect images, we use spatial pyramid model at multi-scale levels, which achieves better robustness to noise and clutter, and thus better copes with severe variations in the pose, scale or rotation. In this paper, we use the 3-scale level pyramid to represent insect image. At the top level of the pyramid, there are 4 image patches which represent the whole image. Each image block size is 50×50. The middle level contains 16 equal size non-overlapping image patches, for which each image block size is 25×25; and the bottom level has 64 image patches for which each image block size is 12×12. Then, the local image features of the 3-level pyramid are combined to represent the insect appearance. The larger scale level provides better geometric features when the classifying insects undergo large appearance variations, while smaller scale level obtains finer features. Finally，the features from fine to coarse levels across different scales are concatenated together to generate the final feature representation of insect images for insect classification. At the second stage, the multiple-kernel learning approach is adopted to combine multiple-space sparse coding. As different features of insect images contribute differently to the classification of insect species, the multiple-space sparse coding technique can combine multiple features of insect species to enhance the recognition performance. Given positive and negative insect samples, the features are extracted. Local image patches of the samples are then represented by multiple-space sparse coding using the corresponding training dictionary. Finally, the multiple-kernel learning classifier is constructed by learning the multiple-space sparse coding of the negative and positive samples for insect categorization and recognition. To meet the need of practical insect image identification, we collected insect images covering various species across several common field crops including corn, soybean, wheat, and canola. Samples of 35 common pest species found in field crops were collected, such as Pieris rapae (Linnaeus), and Leptocorisa acuta (Thunberg). Experimental results showed that our proposed method performed well on the classification of insect species, and outperformed the state-of-the-art methods of the generic insect recognition. Our method improved the recognition rate by more than 9% compared to other methods for the same data sets. In addition, the proposed method had also a good performance and enhanced the average recognition accuracy by 14.1% for the different data sets. [ABSTRACT FROM AUTHOR]

The real-time and accurate recognition for the organs in tomato is crucial to achieve automated agricultural production, such as automatic harvesting and targeted drug application. Due to the difference of fruit maturity and flower age, and the color of tomato organs varying frequently during the growth period, it is very difficult for the simultaneous detection of different tomato organs with the traditional image segmentation method based on color space. In the meanwhile, because of the non-real-time nature of filters such as SIFT (scale-invariant feature transform) and Haar-like, they cannot be directly applied to the real-time detection of plant organs. The convolution neural networks (CNNs) can automatically extract the low-level image feature and high-level semantics of image, and they have real-time performance on GPU (graphics processing unit) devices. Therefore, inspired by Faster R-CNN, especially YOLOv2, in this paper, we proposed a real-time recognition method of main organs in tomato based on CNNs, and designed a corresponding recognition network model which can predict the object boundary and type only using feature map. In greenhouse, the images of various forms of tomato flower, fruit and stem organs were collected, and the image data set of tomato organs was constructed according to the influence of illumination of the image which was considered during the acquisition process. For the sake of screening the underlying network structure for recognition network, the performance and applicability of several typical CNN-based classification networks were analyzed based on the criterion of model size, statistical separability, classification performance, and computation speed. Inspired by the advantages of these typical networks, a channel wise group convolutional (CWGC) block and a corresponding classification network (CWGCNet) were designed. A sample extension training method was presented to further improve the feature extraction ability of these classification networks. The CWGCNet, Darknet-19 and Inception v2 were selected as candidate infrastructure for recognition network. Subsequently, a CWGC block with dropout layer and 3 full convolution layers were respectively attached to the infrastructure to form the overall recognition architecture. Based on the Microsoft Cognitive Toolkit (CNTK), all CNN-based classification networks and recognition networks were implemented by using of Python, and the relevant experiment was performed on a computer equipped with a Tesla K40c GPU. The results show that, compared with the typical CNN-based classification networks, CWGCNet combines high feature statistical separability and real-time performance. On tomato organ image dataset, using Caltech256 to perform sample extension training can significantly improve the feature extraction ability of the classification networks. Compared with the exponential function, the nonlinear scaling factor in the Sigmoid form makes the recognition networks easier to train. In contrast to the 3 full convolution layers, using CWGC block with dropout as an additional convolution layer to the recognition network CNN infrastructure can dramatically reduce the size of the model, while significantly improve the network recall rate, recognition speed and average precision (AP). The convolution part of CWGCNet and the CWGC block with dropout are used as the final structure of the recognition network. The final recognition network can identify the different maturity and different forms of tomato organs, which gets the AP of 96.52%, 97.85% and 82.62% respectively for flower, fruit and stem. The growth stage and maturity of tomato organs have a certain influence on the recognition accuracy, and especially the flowering flower, full maturity fruit and lower stem have higher recognition accuracy. The final network can recall different forms of tomato organs, and the recall rates of flower, fruit and stem can reach 77.39%, 69.33% and 64.23% separately. And the recognition speed of the final network is 62 fps. Compared with YOLOv2, the recall rate can be improved by 14.03 percentage points, and AP can be improved by 2.51 percentage points. [ABSTRACT FROM AUTHOR]

China is the largest country for cultivating kiwifruit, and Shaanxi Province provides the largest production, which accounts for approximately 70% of the production in China and 33% of the global production. Harvesting kiwifruit in this region relies mainly on manual picking which is labor-intensive. Therefore, the introduction of robotic harvesting is highly desirable and suitable. The fast and effective recognition of kiwifruit in the field under natural scenes is one of the key technologies for robotic harvesting. Recently, the study on kiwifruit recognition has been limited to a single cluster and multi clusters in the field have seldom been considered. In this paper, according to growth characteristics of kiwifruit grown on sturdy support structures, an RGB (red, green, blue) camera was placed around 100 cm underneath the canopy so that kiwifruit clusters could be included in the images. We proposed a kiwifruit image recognition system based on the convolutional neural network (CNN), which has a good robustness avoiding the subjectivity and limitation of the features selection by artificial means. The CNN could be trained end to end, from raw pixels to ultimate categories, and we optimized the critical structure parameters and the training strategy. Ultimately, the network was made up of 1 input layer, 3 convolutional layers, 2 sub-sampling layers, 1 full convolutional layer, and 1 output layer. The CNN architecture was optimized by using batch normalization (BN) method, which normalized the data distribution of the middle layer and the output data, accelerating the training convergence and reducing the training time. Therefore, the BN layers were added after the 1, 3 and 5th convolutional layer (Conv1, Conv3, and Conv5 layer) of the original LeNet network. The size of all convolutional kernels was 5×5, and that of all the sub-sampling layers was 2×2. The feature map numbers of Conv1, Conv3, and Conv5 were 6, 16 and 120, respectively. After manual selection and normalizing, the RGB image of kiwifruit was transferred into a matrix with the size of 32×32 as the input of the network, stochastic gradient descent was used to train our models with mini-batch size of 100 examples, and momentum was set as 0.9. In addition, the CNN took advantages of the part connections, the weight sharing and Max pooling techniques to lower complexity and improve the training performance of the model simultaneously. The network used rectified linear units (ReLU) as activation function, which could greatly accelerate network convergence. The proposed model for training kiwifruit was represented as 32×32-6C-2S-16C-2S-120C-2. Finally, 100 images of kiwifruit in the field (including 5918 fruits) were used to test the model, and the results showed that the recognition ratios of occluded fruit, overlapped fruit, adjacent fruit and separated fruit were 78.97%, 83.11%, 91.01% and 94.78%, respectively. The overall recognition rate of the model reached 89.29%, and it only took 0.27 s in average to recognize a fruit. There was no overlap between the testing samples and the training samples, which indicated that the network had a high generalization performance, and the testing images were captured from 9 a.m. to 5 p.m., which indicated the network had a good robustness to lightness variations. However, some fruits were wrongly detected and undetected, which included the fruits occluded by branches or leaves, overlapped to each other and the ones under extremely strong sunlight. Particularly, 2 or more fruits overlapped were recognized as one fruit, which was the main reason to the success rate not very high. This phenomenon demands a further research. By comparing with the conventional methods, it suggested that the method proposed obtained a higher recognition rate and better speed, and especially it could simultaneously identify multi-cluster kiwifruit in the field, which provided significant support for multi-arm operation of harvesting robotic. It proves that the CNN has a great potential for recognition of fruits in the field. [ABSTRACT FROM AUTHOR]

Observing animal's individual and social behaviors is the most effective way to assess animal welfare and healthy. Automated trajectory tracking based on head/tail locations is supposed to be extremely helpful for the realization of pig behavior recognition, especially for group housed pigs in the commercial pig facility. The methods of trajectory tracking for group housed pigs based on head/tail location were described in this paper. The video of group housed nurseries was taken in a commercial pig breeding farm of Hubei Jinlin Original Breeding Swine Co. Ltd. on January 12th, 2016. A high resolution camera (Woshida CL03) was used to record 15 min video. Afterwards, image frames were extracted from the original video in a one-second time interval. Image frames were processed in a computer (configured with IntelCore i7-4790 CPU (central processing unit), 3.6 GHz, 8 G memory) with MATLAB software platform. The image processing for each image frame included 4 steps: background removal, pigs division, head/tail identification and trajectory tracking modification. The background removal was based on the RGB (red, green, blue) color space, from which a vector of RGB mean values of the pig's body was calculated. If the Euclidean distance between the RGB values of one pixel and the RGB mean values vector was less than a small threshold of 100, the pixel was involved in a pig body area and set as 1. Otherwise, it was outside any pig body area and set as 0. When all pixels of the image frame were scanned and calculated by this method, a binary image was acquired. The white area referred to pig's body area, while the black area referred to the background. After that, the morphology erosion and expansion were utilized before the watershed segmentation algorithm to improve the dividing effect for the pigs with adhesion. Pigs division was implemented on the binary images with the improved watershed segmentation algorithm. To discriminate each pig in each image frame, a video tracking and marking method was necessary to be implemented in the video. After being manually marked with the identify number in the first frame, each pig had a unique number and was labelled automatically throughout the video. Abstracting image frames from the video with a very short time interval (1 s), the distance of 2 centroids of the identical pig between 2 continuous image frames would be sufficiently small. Therefore, the video tracking was to find the pig with the closest distance in the next image frame and mark it with the same identify number of the current pig until all the pigs were marked. After each pig was marked throughout the video, using the head/tail location as the coordinates of the pig, the trajectory of each pig in herd could be tracked by the trajectory calculation. Extracting the outline of each pig in frames, the head and the tail outlines were divided from the whole outline, after a sixth of whole outline distance was moved along the outline in 2 opposite directions from the 2 intersection points of the outline and short axis of the minimum bounding rectangle. After the head/tail outline curve was gained from each pig outline, 2 recognition algorithms, the analogous Hough clustering recognition algorithm and the roundness recognition algorithm, were employed to identify the head/tail of each pig. Thus the location of the pig's head/tail could be spotted by locating the centroid of the heat/tail curve. Then the trajectory tracking of the pigs was calculated based on the location of head/tail, and corrected by the motion trends of pigs. Experiment showed that the background was successfully removed from each image frame using the Euclidean distance of RGB values between the pixels and the mean value vector. The improved watershed segmentation algorithm has been verified as an effective tool to divide the pigs with adhesion. The identify number of each pig was tracked from the first frame to the end. The average recognition rate of analogous Hough clustering algorithm was 71.79% for the identification of pig's heads/tails, while the roundness algorithm was 79.67%, which was less sensitive to the distortion of head outline curve. If not including the pigs outside the camera range, the recognition rates would be up to 75% and 85.7% respectively. The roundness algorithm shows an obvious advantage in comparison. The modified trajectory of each pig shows a high agreement with the manually labelled trajectory. More understanding for pigs' behaviors can be acquired from the trajectory of head/tail locations. This trajectory tracking method provides a good reference for further research of behavior recognition. [ABSTRACT FROM AUTHOR]