2.3.1. Adaptive Neighborhood Radius The premise of accurate pose estimation is to construct a feature descriptor of the pallet point cloud with high efficiency, strong robustness and high accuracy; the neighborhood radius is an important factor affecting the performance of feature descriptors. The neighborhood radius of traditional FPFH is usually set to a fixed value according to experience, which reduces the speed of feature extraction. A neighborhood radius selection criterion based on adaptive neighborhood feature entropy is proposed to obtain the neighborhood radius of each key point q k  (points in the key point set Q k ) adaptively. The detailed steps are as follows: Set the range of point cloud neighborhood search radius r j  from lower limit r min  to upper limit r max  with radius interval r d . The upper and lower limits of the radius range are determined by the average point cloud distance d p , which is defined as follows [ 32 ]: (1) d p = 1 N ∑ d m 
where N represents the total number of points in the pallet point cloud, and d m  represents the distance of each key point q k  from its nearest point. Set  r min = d p ,  r max = 2 d p ,  r j + 1 = r j + Δ r , where  j = 1 , 2 , … . Calculate the covariance matrix and eigenvalues corresponding to different neighborhood radius r j . The neighborhood covariance matrix M is defined as follows: (2) M = [ e 1 e 2 e 3 ] [ λ 1 0 0 0 λ 2 0 0 0 λ 3 ] [ e 1 T e 2 T e 3 T ] 
where  λ 1 , λ 2 , λ 3  are the eigenvalues of the neighborhood covariance matrix, and  e 1 , e 2 , e 3  are the corresponding eigenvectors. According to the eigenvalues, the neighborhood feature entropy function  E ξ  is constructed: (3) E ξ = − ξ 1 ln ( ξ 1 ) − ξ 2 ln ( ξ 2 ) − ξ 3 ln ( ξ 3 ) 
where  ξ i = λ i ∑ λ i , i = 1 , 2 , 3 . The adaptive optimal neighborhood radius r opt  of point cloud is determined based on the minimum criterion of neighborhood feature entropy function, that is, when  E ξ  reaches the minimum value, the corresponding neighborhood radius r j  is the optimal neighborhood radius r opt : (4) r opt   = argmin ( E ξ )

2.3.2. Gaussian Weight-Based Fast Point Feature Histogram The features of traditional FPFH are determined by the neighborhood in the radius r of the query point itself and the neighborhood of its neighborhood points, whose maximum range is 2r. The FPFH neighborhood of a query point is shown in  Figure 3 , where p a1 –p a5  are the neighborhood points of the query point P a  within the neighborhood radius r, and p b –p i  are the neighborhood points of the points p a1 –p a5 . Figure 3 Schematic diagram of FPFH neighborhood. The weight coefficient of the traditional FPFH algorithm is determined only by the Euclidean distance between the query point and the neighborhoods; often, there are huge differences among various weight coefficients. As is shown in  Figure 4 , the feature descriptor of the red query point q r  is mostly influenced by the closest black points q a  and q b , and other points only have little effect, which may cause the feature descriptor to be inaccurate. Figure 4 A neighborhood point diagram of a key point. Therefore, an improved FPFH is proposed; each key point q k  in the key point set Q t  is taken as the query point, and based on the adaptive optimal neighborhood radius r opt , all the neighborhood points q ki  within its neighborhood radius are found. The weight coefficient is redefined with the Gaussian distribution, as follows: (5) w G i = 1 2 π σ exp ( − ( w i − μ ) 2 2 σ 2 ) 
where μ is the average distance of q k  and q ki , and  σ  represents the standard deviation of the distance between q k  and q ki . The weight coefficient variation trend of the original approach and the proposed approach is shown in  Figure 5 . It can be seen that the weight coefficient of the original approach varies greatly, and the weight tends to infinity with the decrease of the distance. The proposed AGWF can avoid the problem that the weight coefficient difference between points is apparently large, and the weight coefficient reaches the maximum when the distance is close to the average value, which reasonably solves the unstable problem of the FPFH descriptor calculation. Figure 5 Weight variation trend of neighborhood points of a key point. Combined with the adaptive optimal neighborhood radius selection criterion, an Adaptive Gaussian Weight-based Fast Point Feature Histogram (AGWF) is proposed. It can not only improve the efficiency of feature extraction but also improve the accuracy and robustness; the specific calculation steps are as follows: For each key point q k  in the key point set Q t , search all the neighborhood points q ki  within its optimal neighborhood radius r opt . Compute the normal vectors n s  and n t  corresponding to q k  and q ki , calculate the relative position deviation  ( α , ϕ , θ )  between n s  and n t , and generate the Simple Point Feature Histograms (SPFH) of q k (SPFH(q k )). Local coordinate system (u,v,w) is defined to calculate this deviation, which is shown in  Figure 6 : Figure 6 Diagram of local coordinate system (u,v,w). The calculation formula of relative deviation is as follows: (6) α = v ⋅ n t φ = u ⋅ ( q k − q k i ) ‖ q k − q k i ‖ 2 θ = arctan ( w ⋅ n t , u ⋅ n t ) 3. Then, search the neighborhood points of q ki  based on the adaptive optimal neighborhood radius r opt , and generate the SPFH of q ki (SPFH(q ki )). Based on the Gaussian weight w Gi , the SPFH(q ki ) is weighted to obtain the Adaptive Gaussian Weight-based Fast Point Feature Histogram of key points q k (AGWF(q k )): (7) A G W F P F H ( q k ) = S P F H ( q k ) + 1 k ∑ i = 1 k w G i S P F H ( q k i )

2.4.1. Coarse Registration The purpose of coarse registration is to obtain the initial pose relationship between the source point cloud P t  and the target point cloud, so as to overcome the shortcomings of the ICP algorithm, which requires high initial pose and is easy to fall into local optimum. The SAC-IA algorithm can effectively adjust the initial pose relationship between the source point cloud P and the target point cloud Q and improve the accuracy of pose estimation. The specific steps are as follows: Compute AGWF feature descriptors for all key points in the source point cloud P and the target point cloud Q. N sample points P u  (u = 1, 2, …, N) are randomly selected from the source point cloud P, and the distance between two sample points is greater than the preset distance threshold d min . According to the AGWF, search the closest points Q u  (u = 1, 2, …, N) in the target point cloud Q to the sample points P u , and obtain the initial match point pairs. Obtain the rigid transformation matrix  M 1  between initial match point pairs by Singular Value Decomposition (SVD). Set a registration threshold e l , and calculate the distance function H(l i ) to evaluate the point cloud registration performance. The expression of the distance function H(l i ) is as follows:
 (8) H ( l i ) = { 1 2 l i 2 ‖ l ‖ i < e l 1 2 e l ( 2 ‖ l ‖ i − e l ) ‖ l ‖ i ≥ e l 
where l is the average Euclidean distance between the source point cloud, and I is the number of iterations. Repeat the above four steps; when H(l i ) reaches the minimum, the corresponding transformation matrix is the coarse registration rigid transformation matrix  M c . The rigid transformation of the source point cloud P is carried out based on  M c  to obtain the point cloud P r , and coarse registration is completed.